Resource scheduling method and system based on multi-dimensional load prediction

Abstract

The present invention discloses a resource scheduling method and system based on multi-dimensional load prediction, which relates to the field of cloud computing and intelligent scheduling technology, and is used to solve the problem of efficient prediction and adaptive scheduling under multi-dimensional resource load fluctuations; by collecting and pre-processing historical resource usage data, a multi-dimensional time series sample is constructed; secondly, future load is predicted, and the influence of abnormal fluctuations is suppressed by combining a mutation detection mechanism; thirdly, task priority is evaluated based on task request characteristics and predicted load results, and a preliminary scheduling plan is generated to achieve reasonable resource allocation; finally, key feedback indicators in the scheduling execution process are collected, a scheduling quality scoring system is constructed, and the scheduling strategy is dynamically adjusted and optimized through a multi-objective optimization algorithm. This method has the characteristics of high prediction accuracy, fast scheduling response, and strong adaptability, and can effectively improve resource utilization and task completion efficiency, and reduce scheduling delays and default risks.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing and intelligent scheduling, and more specifically, to a resource scheduling method and system based on multi-dimensional load prediction. Background Art

[0002] In modern cloud computing platforms, resource scheduling is a key technology for ensuring efficient system operation and service quality. With the diversification of business types and the dynamic changes in user needs, the number of scheduling tasks has increased dramatically. Resource requests are multi-dimensional, high-frequency, and highly sudden, leading to significant fluctuations in system load and decreased resource utilization efficiency. Traditional scheduling methods, which rely on static rules, fixed priorities, or simple resource utilization predictions, are difficult to adapt to complex dynamic scheduling scenarios and can easily lead to uneven resource allocation, task execution delays, and even service defaults.

[0003] At the same time, resource usage behavior exhibits significant temporal and cyclical characteristics, and non-stationary changes can occur during sudden load fluctuations or task type switching, placing higher demands on load forecasting and scheduling strategies. Existing scheduling systems generally lack the ability to deeply model multidimensional historical data, unable to accurately capture future load trends. They also lack effective feedback optimization mechanisms, making it difficult to achieve sustained improvement in scheduling effectiveness.

[0004] In view of the above problems, the present invention proposes a solution. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a resource scheduling method and system based on multi-dimensional load prediction to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In a preferred embodiment, it comprises:

[0008] Step 1: Collect and preprocess historical resource usage data to construct a normalized time series input sample;

[0009] Step 2: Predict future resource loads, identify task switching, request fluctuations, and sudden jumps, and optimize model inputs.

[0010] Step 3: Construct the feature vector, combine it with the predicted load trend, calculate the task priority, and generate a preliminary scheduling plan;

[0011] Step 4: Monitor the actual scheduling effect, adjust the scheduling strategy and perform closed-loop iterative optimization through feedback analysis and multi-objective optimization algorithm.

[0012] In a preferred embodiment, in step 1, historical resource data is extracted, normalized, and then converted into a set of fixed-length input sequence samples to generate a historical resource sequence data set.

[0013] In a preferred embodiment, in step 2, the long-term dependencies and local patterns in the historical resource sequence data set are extracted through the gated recurrent unit and the convolutional neural network, and the resource load prediction value at the next moment or future time point is output.

[0014] In a preferred embodiment, in step 2, the switching frequency of task request types in the input historical resource sequence data set, the irregularity of resource request patterns and the mutation jump identification index of the historical resource sequence data set are monitored in real time, and the mutation degree of the historical resource sequence data set at this time is determined by weighted summation, and a mutation degree threshold of the historical resource sequence data set is set. When the mutation degree of the historical resource sequence data set at this time is greater than or equal to the mutation degree threshold of the historical resource sequence data set, the current time step is marked as a mutation point. When the mutation degree of the historical resource sequence data set at this time is less than the mutation degree threshold of the historical resource sequence data set, no mutation processing operation is required.

[0015] In a preferred embodiment, in step 2, the increment between the current input and the previous moment is calculated, and the increment is compressed using a hyperbolic tangent function for the input change at the mutation point, and then replaced with the original mutation input into the GRU;

[0016] The state difference of the historical resource sequence dataset at this time is calculated, and the state difference threshold ϵ is set. If the state difference of the historical resource sequence dataset at this time is greater than or equal to the state difference threshold ϵ, it is considered that the state of the historical resource sequence dataset at this time has jumped. The output of the gated recurrent unit at this time is clipped by the amplitude threshold, and the resource load prediction value adjustment model is established by combining the state difference and the input increment.

[0017] In a preferred embodiment, in step 3, task request data is received from an upstream scheduling interface and converted into a feature vector that can be used for scheduling optimization.

[0018] In a preferred embodiment, in step 3, a weighted scoring function is used to define the task priority score to form an initial scheduling sort list, the predicted load trend data obtained in step 2 is input, and the scheduling window is set according to the load threshold to form a first resource scheduling plan.

[0019] In a preferred embodiment, in step 4, a scheduling solution is deployed and feedback monitoring is performed, and scheduling optimization is performed based on the feedback results.

[0020] In a preferred embodiment, it includes: a data processing module, a load prediction module, a scheduling decision module and a feedback optimization module, and signal connections between the modules;

[0021] The data processing module is mainly used to collect and preprocess historical load resource data and construct normalized time series input samples;

[0022] The load prediction module is mainly used to predict future resource loads, identify task switching, request fluctuations, and sudden jumps, and optimize model inputs;

[0023] The scheduling decision module constructs feature vectors, combines them with predicted load trends, calculates task priorities, and generates a preliminary scheduling plan;

[0024] The feedback optimization module is mainly used to monitor the actual scheduling effect, adjust the scheduling strategy and perform closed-loop iterative optimization through feedback analysis and multi-objective optimization algorithm.

[0025] The technical effects and advantages of the energy-saving and environmentally friendly rail transit turnout snow melting system of the present invention are as follows:

[0026] The present invention realizes the forward-looking and dynamic coordination of resource management by integrating multi-dimensional load prediction and intelligent scheduling optimization. Compared with traditional methods, this solution has significant technical advantages in the following aspects: First, it can fully utilize historical CPU, memory, bandwidth and other multi-source monitoring data to achieve fine-grained time series modeling; second, the introduction of the GRU-CNN hybrid structure combined with the mutation detection mechanism effectively improves the prediction accuracy and stability in burst load scenarios; third, the scheduling strategy is not only based on the static attributes of the task, but also fully considers the predicted load trend and system resource status to achieve dynamic priority evaluation and reasonable arrangement of tasks; fourth, a closed-loop feedback optimization mechanism is designed to reversely optimize the strategy parameters through the scheduling execution results, and continuously improve the system scheduling efficiency and resource utilization. In summary, the present invention significantly enhances the adaptive ability and service guarantee capability of the resource scheduling system in a complex cloud environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is an optimization flow chart of the resource scheduling method and system based on multi-dimensional load prediction of the present invention.

[0028] Figure 2 This is a flowchart of the implementation of the resource scheduling method and system based on multi-dimensional load prediction of the present invention.

[0029] Figure 3 This is a timing diagram of the resource scheduling method and system based on multi-dimensional load prediction of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] Embodiment: The present invention discloses a resource scheduling method based on multi-dimensional load prediction, such as Figure 3 Shown, including:

[0032] Step 1: Preprocess historical resource usage data and construct the time series input required for prediction.

[0033] First, extract historical resource data from time series logs generated by resource management platforms such as Prometheus and Kubernetes Metrics Server. The extracted fields include:

[0034] CPU utilization (unit: %);

[0035] Memory usage (in MB);

[0036] Bandwidth usage (unit: Mbps);

[0037] Task request frequency (unit: times / minute);

[0038] Among them, all fields must contain a unified timestamp and set a fixed sampling interval. If the original records are discontinuous and there are time gaps, forward fill or linear interpolation methods are used to perform time alignment and missing processing to restore data continuity.

[0039] Afterwards, a multidimensional time series data table is generated based on the extracted historical resource data, which is recorded as Rt;

[0040] Furthermore, we traverse each column of historical resource data and independently calculate its mean μ and standard deviation σ, and use the Z-score standardization method to standardize the corresponding value at each time point, specifically according to the formula: zi = (xi - μ) / σ;

[0041] Among them, xi represents the original value, and zi represents the normalized value;

[0042] Afterwards, for features with small fluctuations or constant values, anomaly detection conditions are set to avoid division by 0, and finally a normalized data table Rt′ is generated.

[0043] Construct a sliding time window sequence to convert the normalized continuous historical data into a set of fixed-length input sequence samples. Specifically, first set the time window length T, which indicates that each input sample contains data from the past T time points, and the prediction target offset step length tf, which indicates that each input sample corresponds to the predicted data at the T+tfth time point. Then, traverse and slice the normalized historical resource data Rt′, generating a pair of samples each time.

[0044] The sliding logic is as follows:

[0045] The first set of samples:

[0046] Input: {R1′, R2′, ..., RT′};

[0047] Output target: RT+tf′;

[0048] The second set of samples:

[0049] Input: {R2′,R3′,...,RT+1′}

[0050] Output target: RT+1+tf−1′;

[0051] And so on, the window slides forward one step each time until the end, and finally generates a historical resource sequence dataset.

[0052] Step 2: Use a deep learning model to train the processed historical resource sequence data to predict future load demand.

[0053] First, based on the GRU-CNN hybrid deep neural network model, the gated recurrent unit (GRU) is used to recursively model the previous and next states of the training dataset generated in step 1, dynamically memorizing key load trends.

[0054] Specifically, the control unit includes:

[0055] Update gate: determines whether to retain past information at the current moment;

[0056] Reset gate: determines the degree of integration between the current input and the past state;

[0057] The calculation process is as follows:

[0058] First calculate the output vector of the update gate , according to the formula: , where xt represents the input vector of the current time step, ht-1 represents the hidden state of the previous time step, Iz represents the weight matrix input to the update gate, and Uz represents the weight matrix from the hidden state to the update gate;

[0059] Calculate the output vector of the reset gate , specifically according to the formula: , where Ir represents the weight matrix input to the reset gate, and Ur represents the weight matrix from the hidden state to the reset gate;

[0060] Afterwards, the candidate hidden states are calculated , according to the formula: , where Ih represents the weight matrix from input to candidate hidden state, Uh represents the weight matrix from hidden state to candidate hidden state, and ⊙ represents element-wise multiplication;

[0061] Finally calculate the final hidden state of the current time step , specifically according to the formula: ,in, Indicates the weight of old information, Indicates the weight of new information.

[0062] Afterwards, each input sequence is mapped into a set of hidden state sequences H = [h1,h2,...,h_T], H∈ℝ^{T×D}, where the output H contains the overall trend and periodic characteristics of the historical load data;

[0063] Furthermore, the GRU output H is used as a one-dimensional "feature map" and multiple 1D convolution kernels of different sizes are applied along the time dimension:

[0064] Ci=ReLU(Wi*H+bi), i=1, 2,...,K;

[0065] Among them, Wi represents the i-th convolution kernel, * represents a one-dimensional convolution operation; K represents the number of convolution kernel groups, which is used for multi-scale feature extraction.

[0066] After that, the convolution result is sent to the maximum pooling layer for dimensionality reduction, retaining the most significant response, according to the formula:

[0067] Pi=max(Ci);

[0068] All Pi are concatenated into a one-dimensional feature vector P = [P1, P2, ..., PK] for final prediction to further improve accuracy and generalization performance.

[0069] After training is completed, let the output vector be P, which is input to one or more layers of fully connected neurons for linear or nonlinear mapping:

[0070] Y^=f(WpP+bp);

[0071] Where WpP represents the weight of the fully connected layer, and f represents the ReLU or linear activation function;

[0072] Then, a single-step prediction of the CPU load at the next moment or a multi-step prediction of the complete load sequence at future time points is performed to form the final resource load prediction value.

[0073] It should be noted that, considering that in a heterogeneous cloud platform environment with high-frequency elastic expansion and contraction for multiple users, there are problems such as sudden changes in task types and a sharp increase in the randomness of resource request patterns. The load state will show high instability, and there will be non-stationary mutation characteristics in the historical resource sequence dataset. The GRU model is difficult to capture the mutation pattern, and at the same time, it causes the model convolution weight to shift in response to the actual load. Therefore, in this embodiment, when learning the long-term dependencies in the historical resource sequence dataset and extracting features through the gated recurrent unit GRU and the CNN, the following three features of the input historical resource sequence dataset are monitored:

[0074] Task request type switching frequency monitoring:

[0075] Specifically, a task type classification index set is established based on the input historical resource sequence dataset, and the number of task type switching times ϕt in each time window [t−Δt,t] is counted;

[0076] Detection of irregularities in resource request patterns:

[0077] The coefficient of variation (CV) is used to calculate the volatility of resource request distribution, according to the formula:

[0078] CVt=σ(Xt) / μ(Xt);

[0079] Identification of sudden jumps in historical load data:

[0080] Analyze the input historical resource sequence data set and calculate the first-order difference sequence according to the formula:

[0081] ΔXt=Xt-X(t-1)

[0082] Furthermore, based on the task request type switching frequency, resource request pattern irregularity, and historical load data mutation jump identification indicators calculated above, the mutation degree Ψt of the historical resource sequence data set at this time is determined through a weighted sum evaluation function, specifically according to the formula:

[0083] Ψt=α1×ϕt+α2×CVt+α3×|||ΔXt|;

[0084] Among them, α1 represents the weight of the switching frequency of task request types, α2 represents the weight of the irregularity of resource request patterns, and α3 represents the weight of the identification of mutation jumps in historical load data. The historical load data mutation degree threshold YΨ is set, and the mutation degree Ψt of the historical resource sequence data set at this time is compared with the historical load data mutation degree threshold YΨ. When the mutation degree Ψt of the historical resource sequence data set at this time is greater than or equal to the historical load data mutation degree threshold YΨ, the current time step is marked as a mutation point. When the mutation degree Ψt of the historical resource sequence data set at this time is less than the historical load data mutation degree threshold YΨ, no mutation processing operation is required.

[0085] When a sudden change is detected, the current operation is interrupted immediately and the increment between the current input and the previous moment is calculated according to the formula:

[0086] Δin=Xt-X(t-1);

[0087] The hyperbolic tangent function is used to compress the input changes of the mutation point and control it within the interval (−1,1). The specific formula is:

[0088] Xt-=X(t-1)+tanh(Δin);

[0089] The compressed input Xt replaces the original mutation input and enters the GRU, converting the mutation point into a gradient point, reducing the fluctuation amplitude of the gate control unit, avoiding excessive stimulation of the gate control unit by the input mutation, and weakening the prediction offset caused by the abnormal point;

[0090] Furthermore, the current GRU hidden state is set to ht, and the exponentially weighted moving average (EWMA) is introduced for smoothing:

[0091] ht-=β×ht+(1-β)×h(t-1);

[0092] Next, the state difference of the historical load data at this time is calculated after the output of the smooth gated recurrent unit, according to the formula:

[0093] Δht=||ht-h(t-1)||;

[0094] And set the state difference threshold ϵ. If the state difference Δht of the historical load data at this time is greater than or equal to the state difference threshold ϵ, it is considered that the state of the historical load data at this time has jumped, and the output of the gated recurrent unit at this time needs to be clipped by the amplitude threshold;

[0095] Furthermore, a resource load forecast adjustment model is established by combining the state difference and input increment:

[0096] Y^t=f(ht′)+λ1×tanh(Δht)+λ2×tanh(Δin);

[0097] Among them, Yt^ represents the final optimized resource load prediction value, f(ht′) represents the initial prediction value output by CNN-GRU, and λ1 and λ2 represent the strength of the control optimization items.

[0098] Step 3: Describe the scheduling request in multiple dimensions to determine its importance and resource requirements, and generate a preliminary scheduling plan based on the predicted value and feature score.

[0099] First, the task request data from the upstream scheduling interface is automatically received through the task scheduling interface REST API or message queue, including: request ID, request timestamp, requested resource amount, task type, request priority and task estimated execution time, etc., and formed into a structured data table format;

[0100] Furthermore, the following technical means are used to perform multi-dimensional feature engineering processing to convert the above structured task information into feature vectors that can be used for scheduling optimization:

[0101] Type vectorization:

[0102] Use One-Hot encoding to map task types into discrete vectors;

[0103] Time normalization:

[0104] Convert submission timestamps to work period factors and normalize to [0,1];

[0105] Normalization of resource request amounts:

[0106] Use the Min-Max normalization method to handle memory and bandwidth in the same way;

[0107] Each task is represented as a standardized feature vector that can be used for scoring and ranking, forming the first scheduling feature set.

[0108] Afterwards, the task priority scores are defined using a weighted scoring function:

[0109] Si = α1 × Ti + α2 × Ri + α3 × Pi;

[0110] Among them, Ti represents the time feature score, Ri represents the resource request feature score, and Pi represents the task priority;

[0111] And sort all tasks from high to low according to the score Si to form the initial scheduling sort list;

[0112] Furthermore, the predicted load trend data obtained in step 2 is input, and a scheduling window is set according to the load threshold;

[0113] After that, tasks are selected one by one from the task list, and the earliest time period that can meet their resource requests and estimated duration is searched in turn. If the current time period load + the task resource consumption is less than the set upper limit, the task is inserted. If the task delay time exceeds its maximum allowable scheduling delay, it is marked as unschedulable. Each scheduled task is recorded, including: task ID, start time, allocated resources and execution duration;

[0114] A first resource scheduling plan is formed, which includes all task information that can be reasonably scheduled under the current predicted load, and clarifies the scheduling order, execution time period and resource allocation quantity of each task.

[0115] Step 4: Deploy the scheduling plan and perform feedback monitoring, and optimize the scheduling based on the feedback results;

[0116] First, the scheduling instructions are translated into resource control instructions, such as container deployment scripts or task orchestration instructions;

[0117] Afterwards, use Kubernetes containers or OpenStack virtual machine scheduling to deploy tasks according to the scheduling plan. Deploy the first resource scheduling solution to the scheduler. At the same time, deploy a resource monitoring agent on each node to collect the following key performance indicators during the task scheduling process:

[0118] Resource utilization;

[0119] Task completion rate and SLA breach rate;

[0120] Energy consumption data;

[0121] Scheduling delay and task response time, etc.

[0122] Structuring the feedback data to form a first feedback result;

[0123] and using a time sliding window to count the feedback data to form a first feedback result data set;

[0124] Furthermore, a comprehensive scheduling quality scoring function is constructed to comprehensively evaluate the first feedback result, wherein the comprehensive scheduling quality scoring function is as follows:

[0125] Q=λ1×U+λ2×(1-D)+λ3×(1-T);

[0126] Where U represents the average resource utilization, D represents the task SLA violation rate, T represents the average scheduling delay, and λi represents the weight coefficient;

[0127] Set a preset target Qth, and compare the score Q of the first feedback result with the preset target Qth. When the score Q of the first feedback result is greater than or equal to the preset target Qth, it indicates that the preset target has been achieved; when the score Q of the first feedback result is less than the preset target Qth, it indicates that the scheduling result has not achieved the preset target and enters the optimization process.

[0128] The optimization process is as follows: Figure 1 As shown:

[0129] Step Y1: Convert the feedback information into a structured indicator matrix, where each row represents a feedback record for a task scheduling cycle and each column represents a key indicator. Then, use correlation analysis such as the Pearson correlation coefficient ρ to identify the degree of influence between each feedback indicator and the overall task completion rate.

[0130] Afterwards, threshold detection and Z-score anomaly detection methods are used to identify performance imbalance points, determine whether the corresponding scheduling policy configuration causes system bottlenecks, and finally output the bottleneck constraint set as optimization input.

[0131] Step Y2: Define a set of objective functions, including the following typical objectives:

[0132] f1: maximize task completion rate;

[0133] f2: minimize total energy consumption;

[0134] f3: minimize scheduling delay;

[0135] f4: Minimize the SLA breach rate;

[0136] Furthermore, an optimization model is constructed, and the multi-objective optimization algorithm NSGA-II is applied to generate candidate scheduling schemes that meet the balance of different scheduling requirements.

[0137] Step Y3: Based on the candidate scheduling solutions generated in step Y2, each objective value is processed using a normalization method to obtain a standardized evaluation matrix, and the objective function is adjusted in direction to convert all objectives into a minimization form;

[0138] Furthermore, a weight wi is assigned to each target to satisfy ∑wi=1, and the comprehensive score of each candidate solution is calculated; then, the optimal solution is selected to replace the original scheduling solution based on the comprehensive score of each candidate solution to form the first scheduling optimization solution. The optimized solution replaces the original scheduling solution and is redeployed to enter the next scheduling cycle, forming a closed-loop scheduling system of prediction-execution-feedback-optimization.

[0139] The present invention also proposes a resource scheduling system based on multi-dimensional load prediction, such as Figure 2As shown, it includes: data processing module, load prediction module, scheduling decision module and feedback optimization module, and signal connections between each module;

[0140] The data processing module is mainly used to collect and preprocess historical resource usage data and construct normalized time series input samples;

[0141] The load prediction module is mainly used to predict future resource loads, identify task switching, request fluctuations, and sudden jumps, and optimize model inputs;

[0142] The scheduling decision module constructs feature vectors, combines them with predicted load trends, calculates task priorities, and generates a preliminary scheduling plan;

[0143] The feedback optimization module monitors the actual scheduling effect, adjusts the scheduling strategy and performs closed-loop iterative optimization through feedback analysis and multi-objective optimization algorithms.

[0144] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0145] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0146] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application of the technical solution and the invention constraints. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0147] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0148] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0149] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A resource scheduling method based on multi-dimensional load prediction, characterized in that: include: Step 1: Collect and preprocess historical resource usage data to construct a normalized time series input sample; Step 2: Predict future resource loads, identify task switching, request fluctuations, and sudden jumps, and optimize model inputs. Step 3: Construct the feature vector, combine it with the predicted load trend, calculate the task priority, and generate a preliminary scheduling plan; Step 4: Monitor the actual scheduling effect, adjust the scheduling strategy through feedback analysis and multi-objective optimization algorithm, and perform closed-loop iterative optimization; In step 2, the switching frequency of task request types, the irregularity of resource request patterns, and the mutation jump identification index of the historical resource sequence dataset in the input historical resource sequence dataset are monitored in real time, and the mutation degree of the historical resource sequence dataset at this time is determined by weighted summation. A mutation degree threshold of the historical resource sequence dataset is set. When the mutation degree of the historical resource sequence dataset at this time is greater than or equal to the mutation degree threshold of the historical resource sequence dataset, the current time step is marked as a mutation point. When the mutation degree of the historical resource sequence dataset at this time is less than the mutation degree threshold of the historical resource sequence dataset, no mutation processing operation is required. In step 2, the increment between the current input historical resource sequence dataset and the previous moment’s historical resource sequence dataset is calculated, and the hyperbolic tangent function is used to compress the increment of the historical resource sequence dataset for the input change of the mutation point, and the increment is input into the GRU instead of the original mutation. The state difference of the historical resource sequence dataset at this time is calculated, and the state difference threshold ϵ is set. If the state difference of the historical resource sequence dataset at this time is greater than or equal to the state difference threshold ϵ, it is considered that the state of the historical resource sequence dataset at this time has jumped. The output of the gated recurrent unit at this time is clipped by the amplitude threshold, and the resource load prediction value adjustment model is established by combining the state difference and the input increment.

2. The resource scheduling method based on multi-dimensional load prediction according to claim 1, characterized in that: In step 1, historical load resource data is extracted, normalized, and then converted into a set of fixed-length input sequence samples to generate a historical resource sequence dataset.

3. The resource scheduling method based on multi-dimensional load prediction according to claim 2, characterized in that: In step 2, the long-term dependencies and local patterns in the historical resource sequence dataset are extracted through the gated recurrent unit and convolutional neural network, and the resource load prediction value at the next moment or future time point is output.

4. The resource scheduling method based on multi-dimensional load prediction according to claim 1, characterized in that: In step 3, the task request data from the upstream scheduling interface is received and converted into a feature vector that can be used for scheduling optimization.

5. The resource scheduling method based on multi-dimensional load prediction according to claim 4, characterized in that: In step 3, a weighted scoring function is used to define the task priority score to form an initial scheduling sort list, the predicted load trend data obtained in step 2 is input, and the scheduling window is set according to the load threshold to form a first resource scheduling plan.

6. The resource scheduling method based on multi-dimensional load prediction according to claim 1 is characterized in that ; In step 4, deploy the scheduling plan and perform feedback monitoring, and perform scheduling optimization based on the feedback results.

Images