A spatiotemporal prediction-based elastic computing power scheduling method, device and medium

Abstract

The application provides a kind of elasticity computing power scheduling method, device and medium based on space-time prediction, belong to cloud computing and edge computing technical field.The method includes: real-time collection historical load, satellite cloud picture, meteorological data, sudden traffic mark, energy price, green electricity proportion;Build fusion prediction model, including multi-modal feature encoding module, space-time attention fusion module and prediction output layer;Build multi-objective optimization function, minimize cost, delay, resource fluctuation and maximize green reward;Task allocation variables and resource adjustment amount are solved by using mixed integer programming;Monitor actual green electricity utilization rate, dynamically adjust green reward coefficient to correct prediction and scheduling deviation.The application can predict the computing power demand fluctuation of different regions in the future period in advance, so as to dynamically adjust resource allocation, reduce resource idling and waste, significantly improve the utilization rate of computing power resources.

Description

TECHNICAL FIELD

[0001] The application relates to a spatiotemporal prediction-based elastic computing power scheduling method and device and medium, and belongs to the technical field of cloud computing and edge computing. BACKGROUND

[0002] With the rapid development of cloud computing and edge computing technologies, the computing power demand of data centers, edge nodes, and hybrid clouds is increasing, and higher requirements are put forward for efficient use and dynamic adaptation of resources. Traditional computing power network management solutions often focus on static resource allocation, which is difficult to adapt to complex and changing application requirements and environmental changes, resulting in low resource utilization, high operating costs, and difficulty in guaranteeing service quality.

[0003] In the prior art, for example, comparative document CN119583557A proposes a computing power network management scheme based on energy consumption-performance double index fusion. This scheme realizes preliminary optimization of the computing power network through steps such as business request analysis, routing fusion modeling, node screening mechanism, path scoring system, and digital twin verification. However, this scheme still has the following shortcomings:

[0004] Insufficient prediction mechanism: the existing technology mainly uses static historical data analysis models, lacking the ability to jointly predict future computing power demand in space and time. This static analysis method is difficult to accurately capture the dynamic changes of computing power demand, resulting in high resource pre-allocation error rate and inability to effectively respond to sudden demand or long-term trend changes.

[0005] Lack of spatiotemporal correlation: in cross-regional business scenarios, the existing technology does not establish a spatiotemporal coupling model of regional computing power demand. Time zone differences, holiday migration, and other spatiotemporal factors can significantly affect computing power demand, but the existing technology is difficult to effectively predict these fluctuations, resulting in unreasonable resource allocation and affecting the overall performance of the system.

[0006] Green energy adaptation lag: with the continuous development of new energy technologies, green energy such as photovoltaic power generation and wind power is increasingly widely used in computing power networks. However, the dynamic matching mechanism of the existing technology responds slowly, lacks real-time modeling capability for green energy such as photovoltaic power generation and wind power, and cannot fully utilize green energy to reduce operating costs and environmental impact.

[0007] This background section outlines the shortcomings of the existing technology and leads to the technical problems solved by the present application and its innovative points, laying the foundation for subsequent technical solution descriptions. SUMMARY

[0008] The application aims to provide a spatiotemporal prediction-based elastic computing power scheduling method, device and medium, which aims to predict future regionalized computing power demand through a deep spatiotemporal network, dynamically adjust computing power resource allocation in combination with a multi-objective optimization algorithm, and realize the collaborative optimization of computing power resource utilization rate improvement, energy consumption dynamic adaptation, reasonable operation cost control and comprehensive service quality guarantee.

[0009] To achieve the above-mentioned purpose, the application is implemented through the following technical solutions:

[0010] A spatiotemporal prediction-based elastic computing power scheduling method, comprising the following steps:

[0011] Real-time collection of historical load, satellite cloud map, meteorological data, sudden traffic marker, energy price and green electricity proportion;

[0012] A fusion prediction model is constructed, including a multi-modal feature encoding module, a spatiotemporal attention fusion module and a prediction output layer; the multi-modal feature encoding module respectively encodes the satellite cloud map, meteorological data, light intensity, sudden traffic marker and historical load; the spatiotemporal attention fusion module dynamically weights and fuses the encoded features; the prediction output layer outputs a load prediction value based on the fused features, and introduces an explainable compensation term of energy price and green electricity proportion;

[0013] A multi-objective optimization function is constructed to minimize cost, delay and resource fluctuation and maximize green reward;

[0014] A mixed integer programming is adopted to solve task allocation variables and resource adjustment amount;

[0015] The actual green electricity utilization rate is monitored, and the green reward coefficient is dynamically adjusted to correct the prediction and scheduling deviation.

[0016] Preferably, the historical load includes node computing power utilization rate and task queue length, the satellite cloud map includes regional cloud coverage, the meteorological data includes temperature, humidity, wind speed and light intensity, the energy price includes time-of-use electricity price and carbon emission cost, and the green electricity proportion includes renewable energy power supply proportion of each node.

[0017] Preferably, the multi-modal feature encoding module comprises:

[0018] Satellite cloud map encoding: ResNet-18 is adopted to extract spatial features, and a cloud layer movement trend feature is generated through a spatiotemporal convolution layer;

[0019] Meteorological time series encoding: temperature, humidity and wind speed time series features are captured through a bidirectional long short-term memory network, and light intensity fluctuation is modeled through a time convolution network;

[0020] Sudden traffic marker encoding: diffusion effect features are generated based on event labels and a Gaussian decay function;

[0021] Historical load encoding: encode the historical load sequence using a temporal convolution network to extract multi-scale time series features.

[0022] Preferably, the spatio-temporal attention fusion formula of the spatio-temporal attention fusion module is as follows:

[0023] ,

[0024] ,

[0025] ,

[0026] ,

[0027] wherein, is a learnable linear transformation layer that projects features from different sources to the same dimensional space, is the current task requirement, is the satellite cloud image, meteorological data, and sudden traffic marker information feature label, is the specific numerical value of the feature label, is the multi-scale time series feature extracted after encoding the historical load, is the spatio-temporal encoding feature of the satellite cloud image, is the meteorological time series encoding feature, is the sudden traffic marker encoding feature; represents the calculated task requirement vector and the matching degree, represents the square root of the feature dimension length, is a softmax function, represents the fused feature vector.

[0028] Preferably, the specific formula of the prediction output layer is as follows:

[0029] ,

[0030] wherein, represents a temporal convolution network, is an explainability compensation term, represents the influence strength of external factors, represents a hyperbolic tangent function, , represents a learnable weight matrix, represents the energy price, represents the green electricity proportion, represents the predicted load.

[0031] Preferably, the multi-objective optimization function is as follows:

[0032] ,

[0033] ,

[0034] ,

[0035] ,

[0036] wherein, is a task scheduling cost term, denotes a task whether assigned to a node , denotes a unit task cost of a node , denotes a total number of tasks, denotes a total number of nodes;

[0037] is a quality of service cost term, denotes a task expected delay of a task at a node, is a delay weight coefficient;

[0038] is a green scheduling reward term, denotes a green weight of a node , is a green reward intensity coefficient;

[0039] is an adjustment overhead cost term, is a resource elasticity adjustment amount of a node , is a stability coefficient;

[0040] is a task integrity constraint;

[0041] is a capacity constraint, is a maximum physical load of a node , is an elasticity scaling proportional adjustment amount;

[0042] is a green constraint, is a green electricity proportion of a node , is a total number of tasks assigned to a node , is a minimum green electricity proportion threshold.

[0043] The green weight calculation method is preferably as follows:

[0044]

[0045]

[0046] wherein, is the light intensity gain coefficient, represents the theoretical maximum light intensity value, represents the real-time light intensity, is the smoothing term.

[0047] The dynamic adjustment of the green reward coefficient to correct the prediction and scheduling deviation is preferably as follows:

[0048] Real-time calculation of green electricity utilization rate

[0049]

[0050] According to the calculated green electricity utilization rate Dynamic adjustment of the green reward intensity coefficient

[0051]

[0052] wherein, represents the adjusted green reward intensity coefficient, is the adjustment coefficient,

[0053] The adaptive adjustment rule is as follows:

[0054] When,

[0055] When,

[0056] When,

[0057] When the satellite monitors that the cloud cover in a certain area thickens, the prediction output layer is down-regulated , the green weight of the area is reduced, and the green reward intensity coefficient adjustment compensation task migration is triggered.

[0058] An elastic computing power scheduling system based on space-time prediction, comprising:

[0059] ​​​​​​​​​​A data collection module is configured to collect multi-source data in real time, including historical load, satellite cloud images, weather data, sudden traffic markers, energy prices, and green electricity ratios.

[0060] A spatio-temporal prediction module includes a multi-modal feature encoding module, a spatio-temporal attention fusion module, and a prediction output layer. The multi-modal feature encoding module encodes satellite cloud images, weather data, light intensity, sudden traffic markers, and historical load. The spatio-temporal attention fusion module dynamically weights and fuses the encoded features. The prediction output layer outputs load prediction values based on the fused features and introduces an interpretable compensation term for energy prices and green electricity ratios.

[0061] An elastic scheduling module constructs a multi-objective optimization function to minimize cost, delay, and resource fluctuations and maximize green rewards. A mixed integer programming is used to solve task allocation variables and resource adjustment amounts.

[0062] An execution feedback module monitors actual green electricity utilization and dynamically adjusts the green reward coefficient to correct prediction and scheduling deviations.

[0063] An elastic computing power scheduling method and device based on spatio-temporal prediction, including a processor and a memory storing program instructions, the processor is configured to execute the program instructions when running, to execute the elastic computing power scheduling method based on spatio-temporal prediction.

[0064] The advantages of the present application are:

[0065] The system can predict the computing power demand fluctuations in different regions in the future period in advance, dynamically adjust resource allocation, reduce resource idling and waste, and significantly improve the utilization rate of computing power resources.

[0066] The system combines a multi-objective optimization algorithm and comprehensively considers task scheduling cost, service quality cost, green scheduling reward items, and adjustment cost items to generate an optimal computing power resource allocation strategy. This helps to reduce overall operating costs while ensuring service quality.

[0067] The elastic scheduling decision is based on a dynamic scheduling algorithm of a multi-objective optimization equation, which can respond to changes in computing power demand in real time and achieve rapid expansion and contraction of computing power resources and task scheduling. This flexibility enables the system to better adapt to complex and changing application requirements and environmental changes.

[0068] By combining spatio-temporal factors for prediction and scheduling, the system can optimize load balancing between regions, reduce performance bottlenecks caused by high computing power demand in some regions, and improve the utilization rate of computing power resources in other regions.

[0069] The system considers the dynamic adaptation of green energy, preferentially schedules nodes in regions with high green energy proportion and high illumination, thereby fully utilizing green energy, reducing carbon emissions, and improving the environmental contribution of the system.

[0070] The closed-loop feedback optimization module realizes dynamic optimization of green efficiency optimization, dynamically adjusts according to the monitored green energy utilization rate in actual operation, and ensures that the system always remains in an optimal operating state.

[0071] Through the innovative design of a three-level fusion prediction model (including multi-modal feature encoding, spatio-temporal attention fusion, and prediction output layer), the system can more accurately predict future computing power demand and generate more reasonable scheduling strategies, thereby improving the prediction and scheduling accuracy of the overall system. BRIEF DESCRIPTION OF DRAWINGS

[0072] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application.

[0073] Figure 1 The present application is a method flowchart.

[0074] Figure 2 The present application is a system structure schematic diagram.

[0075] Figure 3 The present application is a system module structure schematic diagram. DETAILED DESCRIPTION

[0076] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0077] Embodiment 1

[0078] As shown in the accompanying drawings, an elastic computing power scheduling method based on spatio-temporal prediction includes the following steps: Figure 1

[0079] S1: Real-time collection of historical load, satellite cloud image, meteorological data, sudden traffic marker, energy price, and green energy proportion;

[0080] ​S2: Construct a fusion prediction model, including a multi-modal feature encoding module, a spatio-temporal attention fusion module, and a prediction output layer; the multi-modal feature encoding module encodes satellite cloud images, meteorological data, light intensity, sudden traffic labels, and historical load; the spatio-temporal attention fusion module dynamically weights and fuses the encoded features; the prediction output layer outputs load prediction values based on the fused features, and introduces an interpretable compensation term for energy prices and green electricity proportion;

[0081] S3: Construct a multi-objective optimization function to minimize cost, delay, and resource fluctuation, and maximize green reward;

[0082] S4: Use mixed integer programming to solve task allocation variables and resource adjustment amounts;

[0083] S5: Monitor the actual green electricity utilization rate, and dynamically adjust the green reward coefficient to correct the prediction and scheduling deviation.

[0084] As a refinement of the above embodiment, the historical load includes the computing power usage rate and task queue length of each node, the satellite cloud image includes the regional cloud coverage, the meteorological data includes temperature, humidity, wind speed, and light intensity, the energy price includes time-of-use electricity price and carbon emission cost, and the green electricity proportion includes the proportion of renewable energy power supply of each node.

[0085] As a refinement of the above embodiment, the multi-modal feature encoding module includes:

[0086] (1) Satellite cloud image encoding: ResNet-18 is used to extract spatial features, and a spatio-temporal convolution layer is used to generate cloud layer movement trend features:

[0087] ,

[0088] wherein, is the satellite cloud image taken at the current time,

[0089] is a series of consecutive satellite cloud images taken from the past time to the current time ,

[0090] represents feature concatenation, the spatial features of the current time (the result of ResNet-18) and the change trend in the recent period (the result of ST-Conv) are combined into a more comprehensive feature vector,

[0091] represents a spatio-temporal convolution layer.

[0092] ​(2) Weather time series encoding: Capture temperature, humidity, wind speed time series features through bidirectional long short-term memory network, model light intensity fluctuation through time convolution network:

[0093]

[0094] Refers to the humidity data from the past moment until the current moment .

[0095] Refers to the humidity data from the past moment until the current moment .

[0096] Refers to the wind speed data from the past moment until the current moment .

[0097] Refers to the light data from the past moment until the current moment .

[0098] BiLSTM role:

[0099] Forward LSTM: Learn sequence trends from old to new (e.g. temperature gradually rising).

[0100] Reverse LSTM: Learn sequence dependencies from new to old (e.g. current wind speed is affected by the previous moment).

[0101] Output: Weather feature vector integrating historical context.

[0102] TCN role: Capture local fluctuation patterns through multiple convolution kernels (e.g. light intensity surges during the day / drops in the evening), output light feature vector. If a single LSTM is used to process all weather parameters, the LSTM's memory gate mechanism will smooth out sudden signals, causing predicted values to lag behind true values. By using TCN to handle light fluctuations, the prediction response delay can be effectively reduced.

[0103] (3) Burst traffic label encoding: Based on the Gaussian decay function to quantify the time diffusion effect of burst events, where controls the duration of event impact, when =3 represents that the event impact lasts about ±6 hours.

[0104] ,

[0105] is a certain time point When an important event occurs (such as "Double 11 0 second kill starts"), this label is assigned a specific value. This label itself tells the system that an event has occurred that may cause a traffic surge.

[0106] MLP is a small artificial intelligence model (neural network). Its role is similar to an impact evaluator. It receives the event label as input. After internal calculation, it outputs a value that better represents the intensity of the traffic impact that the specific event may cause.

[0107] is a time decay factor - Gaussian bell curve combined formula, which can simulate the process of event influence gradually weakening over time. represents the current time, represents the specific time when the event occurs, the moment the event occurs ( ), the impact is the largest. The farther away from the event occurrence time point, the smaller the impact. The parameter determines whether the impact decays fast or slow, The larger the value, the flatter the bell curve, the slower the decay rate. The event impact will last longer. Like a long aftereffect. The smaller the value, the sharper the bell curve, the faster the decay rate, and the event impact disappears quickly.

[0108] represents the actual impact value of the burst event on computing power demand at the current time , considering time decay.

[0109] (4) Historical load encoding: encode the historical load sequence using a time convolution network to extract multi-scale time sequence features:

[0110]

[0111] represents the historical load record of each computing node in the time period before the current time , which records the real-time load and task queue of each computing node in the past hours (or minutes, depending on the data collection interval). These data points form a curve that fluctuates over time.

[0112] (time convolution network): along the time axis (from the past hours to the present ) sliding, carefully scanning each load curve, specifically looking for key patterns (short-term fluctuations, medium-term trends, long-term periodicity, dependencies) that exist in the data, by analyzing these complex patterns, Finally output a relatively short but information-rich feature vector This vector is no longer the original load value, but the core information of the load change rule, which is one of the important inputs of the subsequent prediction model.

[0113] As a refinement of the above embodiment, the spatio-temporal attention fusion formula is as follows:

[0114] ,

[0115] ,

[0116] ,

[0117] ,

[0118] Among them, is a learnable linear transformation layer that projects features from different sources into the same dimensional space, is the current task requirement, is the satellite cloud image, weather data, and burst traffic label information feature label, which represents the environmental conditions in which the data center is located, is the specific numerical value of the feature label, is the multi-scale time series feature extracted after encoding the historical load, is the satellite cloud image spatio-temporal encoding feature, is the weather time series encoding feature, is the burst traffic label encoding feature; represents the calculation task requirement vector and the matching degree, the higher the matching degree, the greater the influence of this environmental factor on the current task requirement, represents the square root of the length of the feature dimension, the purpose is to stabilize the gradient and prevent over-saturation, is the softmax function, Softmax converts all the matching degree scores calculated above into a probability distribution (sum = 1). Environmental factors with high scores will get higher weights (attention), and those with low scores will get lower weights, The fused feature vector is represented, and the calculated attention weight is applied to the specific environmental details (V). For each environmental factor, its importance score is used to weight its specific value (V). It is no longer a simple splicing of the original data, but a new feature vector that has fused all environmental information after attention screening, containing the information: "According to the current task load, which environmental factors (satellite cloud map, weather, emergency) and their specific details (numerical values) need to be considered in order to more accurately predict future computing power demand."

[0119] As a refinement of the above embodiment, the prediction output layer is specifically formulated as follows:

[0120] ,

[0121] , wherein, represents a time convolution network, which is the main prediction model, and predicts the basic value of future load through historical rules, is an explainability compensation term, represents the influence strength of external factors, represents a hyperbolic tangent function, which limits the result to the range of [-1, 1] to prevent overcompensation, (negative correlation), (positive correlation) represents a learnable weight matrix, represents the energy price, represents the proportion of green electricity, represents the predicted load, represents the influence of energy price, the higher the energy price, the greater the tendency of the scheduling system to use less computing power, represents the influence of the proportion of green electricity, the higher the proportion of green energy,

[0122] Preferably, the multi-objective optimization function is as follows:

[0123] ,

[0124] ,

[0125] ,

[0126] ,

[0127] wherein, is a task scheduling cost term, and the technical essence is to calculate the total economic expenditure of all task scheduling, represents whether the task is allocated to the node , representing the unit task cost of the node , representing the total number of tasks, representing the total number of nodes;

[0128] is a service quality cost term, the technical essence of which is to quantify the loss of user-perceived service quality, representing the task expected delay of the node , is a delay weight coefficient;

[0129] is a green scheduling reward term, the technical essence of which is to encourage the allocation of tasks to nodes with sufficient green energy, representing the green weight of the node , is a green reward intensity coefficient;

[0130] is an adjustment overhead cost term, the technical essence of which is to prevent service jitter caused by drastic fluctuations in resources, is the resource elasticity adjustment amount of the node , is a stability coefficient;

[0131] is a task integrity constraint, the technical essence of which is to ensure that each task is processed, and is used as a hard constraint in the optimization algorithm;

[0132] is a capacity constraint, the technical essence of which is that the total amount of tasks allocated to the node should be less than the maximum physical load of the node, is the maximum physical load of the node , is an elasticity scaling ratio adjustment amount;

[0133] is a green constraint, the technical essence of which is that the average value of the green energy proportion of the nodes where the tasks allocated by the entire system are located cannot be lower than , is the green electricity proportion of the node , is the total number of tasks allocated to the node , is the minimum green electricity proportion threshold.

[0134] The objective of the multi-objective function is to find an optimal allocation scheme that simultaneously achieves the best balance of the following four objectives:

[0135] The lower the economic cost, the better (less money spent);

[0136] The lower the latency, the better (for user satisfaction);

[0137] The greater the contribution to environmental protection, the better (use more green electricity);

[0138] The smaller the system fluctuation, the better (avoid frequent adjustments).

[0139] Specifically, the green weight is calculated as follows:

[0140] ,

[0141] ,

[0142] in, This is the light intensity gain coefficient. , This represents the theoretical maximum light intensity value. Indicates real-time light intensity. The smoothing term is 0.01. , .

[0143] Nodes in areas with high green electricity ratios and high solar radiation levels should be prioritized for scheduling.

[0144] As a refinement of the above embodiments, the mixed-integer programming solution is as follows:

[0145] Mixed Integer Programming (MIP) is adopted as the core solution method for the elastic scheduling algorithm. Its purpose is to efficiently handle multi-objective optimization problems in computing power scheduling, while simultaneously satisfying the discrete constraints of task allocation and the continuous requirements of resource adjustment. There are two types of key decision variables in computing power scheduling:

[0146] Discrete variables (0-1 decision): Task Assign to a node Such variables naturally require integer solutions (i.e., the task cannot be indivisible).

[0147] Continuous variable: Resource elasticity adjustment Represents a node The scaling up or down of computing power (e.g., adding +3.5 virtual machines).

[0148] Implement using the branch and bound method after relaxing continuous variables:

[0149] Branching: Split the discrete variables into 0 / 1 and enumerate all possible task assignments.

[0150] Bounding: Optimize the objective function value of the current branch in a continuous space.

[0151] Pruning: discard branches whose target values are worse than known feasible solutions, greatly reducing the amount of computation.

[0152] Ensure that the generated scheduling scheme not only meets the physical constraints of the actual system, but also maximizes the overall efficiency of the system.

[0153] As a refinement of the above embodiment, according to the green electricity utilization rate monitored in actual operation Dynamic adjustment, this value may be lower than the theoretical green electricity utilization rate required when solving MIP due to prediction errors, execution deviations or environmental changes, etc.

[0154] Specifically, the dynamic adjustment of the green reward coefficient to correct the prediction and scheduling deviation is as follows:

[0155] Real-time calculation of green electricity utilization rate :

[0156] ,

[0157] According to the calculation of green electricity utilization rate Dynamic adjustment of green reward intensity coefficient :

[0158] ,

[0159] Wherein, represents the adjusted green reward intensity coefficient, is the adjustment coefficient,

[0160] The adaptive adjustment rule is as follows:

[0161] When, ,

[0162] When, ,

[0163] When, ,

[0164] When the satellite monitors that the cloud layer in a certain area thickens, the output layer is down-regulated , reduce the green weight of the area, trigger the green reward intensity coefficient adjustment compensation task migration.

[0165] Embodiment 2

[0166] As Figures 2-3 shown, an elastic computing power scheduling system based on space-time prediction includes:

[0167] a data collection module for collecting multi-source data in real time, including historical load, satellite cloud image, meteorological data, sudden traffic mark, energy price, and green electricity proportion;

[0168] a spatio-temporal prediction module including a multi-modal feature encoding module, a spatio-temporal attention fusion module, and a prediction output layer; the multi-modal feature encoding module encodes satellite cloud image, meteorological data, light intensity, sudden traffic mark, and historical load respectively; the spatio-temporal attention fusion module dynamically weights and fuses the encoded features; the prediction output layer outputs a load prediction value based on the fused features, and introduces an interpretable compensation term of energy price and green electricity proportion;

[0169] an elastic scheduling module for constructing a multi-objective optimization function to minimize cost, delay, and resource fluctuation and maximize green reward; and a mixed integer programming is used to solve task allocation variables and resource adjustment amount;

[0170] an execution feedback module for monitoring actual green electricity utilization rate and dynamically adjusting a green reward coefficient to correct prediction and scheduling deviation.

[0171] The embodiment of the disclosure also provides an elastic computing power scheduling device based on spatio-temporal prediction, including a processor and a memory. Optionally, the device can also include a communication interface and a bus. The processor, the communication interface, and the memory can communicate with each other through the bus. The communication interface can be used for information transmission. The processor can call the logic instructions in the memory to execute the elastic computing power scheduling method based on spatio-temporal prediction of the above-mentioned embodiment.

[0172] In addition, the logic instructions in the memory described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium.

[0173] The memory, as a computer readable storage medium, can be used to store software programs, computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the disclosure. The processor executes the program instructions / modules stored in the memory, thereby executing function applications and data processing, that is, implementing the elastic computing power scheduling method based on spatio-temporal prediction in the above-mentioned embodiment.

[0174] The memory can include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; the data storage area can store data created according to the use of the terminal device, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory.

[0175] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that the technical solutions described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalent ones. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A spatiotemporal prediction-based elastic computing power scheduling method, characterized in that, The method comprises the following steps: Real-time collection of historical load, satellite cloud image, meteorological data, sudden traffic label, energy price, and green electricity proportion; the historical load comprises the computing power utilization rate and task queue length of each node, the satellite cloud image comprises the regional cloud coverage, the meteorological data comprises temperature, humidity, wind speed, and light intensity, the energy price comprises time-of-use electricity price and carbon emission cost, and the green electricity proportion comprises the proportion of renewable energy power supply of each node; A fusion prediction model is constructed, comprising a multi-modal feature encoding module, a spatio-temporal attention fusion module, and a prediction output layer; the multi-modal feature encoding module respectively encodes the satellite cloud image, meteorological data, light intensity, sudden traffic label, and historical load; the spatio-temporal attention fusion module dynamically weights and fuses the encoded features; and the prediction output layer outputs a load prediction value based on the fused features and introduces an explainable compensation term of energy price and green electricity proportion; A multi-objective optimization function is constructed to minimize the task scheduling cost, service quality cost, and adjustment overhead cost and maximize the green scheduling reward; Solving using mixed integer programming and ; The actual green electricity utilization rate is monitored, and a green reward intensity coefficient is dynamically adjusted to correct the prediction and scheduling deviation. The multi-objective optimization function is as follows: ， ， ， ， wherein, is a task scheduling cost term, denotes a task whether assigned to a node , denotes a node unit task cost, denotes the total number of tasks, denotes the total number of nodes, denotes the predicted load; for a quality of service cost item, representing a task at a node an expected delay, is a delay weight coefficient; a green dispatch reward term, a node a green weight, a green reward intensity coefficient; to adjust the cost item of overhead, to the node resource elasticity adjustment amount, stability coefficient; is a task integrity constraint; a capacity constraint, a maximum physical load of a node a resource elasticity adjustment amount of a node a maximum physical load of a node a resource elasticity adjustment amount of a node is a green constraint, is a node green electricity ratio, is a total number of tasks allocated to the node on, is a minimum green electricity ratio threshold.

2. The spatio-temporal prediction based elastic computing power scheduling method of claim 1, wherein, The multi-modal feature encoding module comprises: Satellite cloud image encoding: ResNet-18 is used to extract spatial features, and a spatio-temporal convolution layer is used to generate cloud layer movement trend features; Meteorological time series encoding: a bidirectional long short-term memory network is used to capture temperature, humidity, and wind speed time series features, and a time convolution network is used to model light intensity fluctuations; Sudden traffic label encoding: an event label and a Gaussian decay function are used to generate diffusion effect features; Historical load encoding: a time convolution network is used to encode the historical load sequence to extract multi-scale time series features.

3. The spatio-temporal prediction based elastic computing power scheduling method of claim 2, wherein, The spatio-temporal attention fusion formula of the spatio-temporal attention fusion module is as follows: ， ， wherein, is a learnable linear transformation layer that projects features from different sources into the same dimensional space, is the current task requirement, is the satellite cloud image, weather data, burst traffic label information feature label, and the feature vector is obtained by linear transformation on , is the specific numerical value of the feature label, and the feature vector is obtained by linear transformation on , is the multi-scale time sequence feature extracted after encoding the historical load, is the satellite cloud image spatiotemporal encoding feature, is the weather time sequence encoding feature, is the burst traffic label encoding feature; represents the calculation task requirement vector and the matching degree, represents the square root of the feature dimension length, is a softmax function, represents the fused feature vector.

4. The spatio-temporal prediction based elastic computing power scheduling method of claim 3, wherein, The specific formula of the prediction output layer is as follows: ， wherein, denotes a temporal convolutional network, is an explainability compensation term, denotes an external factor influence strength, denotes a hyperbolic tangent function, , denotes a learnable weight matrix, denotes an energy price, denotes a green electricity proportion, denotes a predicted load, denotes a current time instant.

5. The spatio-temporal prediction based elastic computing power scheduling method of claim 1, wherein, The green weight calculation method is as follows: ， ， wherein, is a light intensity gain coefficient, , denotes a theoretical maximum light intensity value, denotes a real-time light intensity, is a smoothing term.

6. The spatio-temporal prediction based elastic computing power scheduling method of claim 1, wherein, The specific method of dynamically adjusting the green reward intensity coefficient to correct the prediction and scheduling deviation is as follows: Real-time calculation of green electricity utilization : ， According to the calculation of green electricity utilization rate Dynamically adjusting green reward intensity coefficient : ， wherein, represents the adjusted green reward intensity coefficient, is an adjustment coefficient, The adaptive adjustment rule is as follows: Time, , time, , Time, , When the satellite monitors that the cloud in a region thickens, the output layer is predicted to be down-regulated , the green weight in the region is reduced, and the green reward intensity coefficient adjustment compensation task migration is triggered.

7. A spatiotemporal prediction-based elastic computing power scheduling system, characterized in that, The method for elastic computing power scheduling based on spatio-temporal prediction according to any one of claims 1-6 comprises: A data acquisition module is configured to acquire multi-source data in real time, including historical load, satellite cloud image, meteorological data, sudden traffic label, energy price, and green electricity proportion; A spatio-temporal prediction module comprises a multi-modal feature encoding module, a spatio-temporal attention fusion module, and a prediction output layer; the multi-modal feature encoding module respectively encodes the satellite cloud image, meteorological data, light intensity, sudden traffic label, and historical load; the spatio-temporal attention fusion module dynamically weights and fuses the encoded features; and the prediction output layer outputs a load prediction value based on the fused features and introduces an explainable compensation term of energy price and green electricity proportion; An elastic scheduling module constructs a multi-objective optimization function to minimize task scheduling cost, quality of service cost, adjustment overhead cost, and maximize green scheduling reward; and adopts mixed integer programming to solve and ; The multi-objective optimization function is as follows: ， ， ， ， wherein, is a task scheduling cost term, denotes a task is assigned to a node , denotes a node unit task cost, denotes the total number of tasks, denotes the total number of nodes, denotes the predicted load; for a quality of service cost item, representing a task at a node expected delay, is a delay weight coefficient; green scheduling reward term, representing a node green weight, green reward intensity coefficient; to adjust the cost item of overhead, to adjust the resource elasticity of the node , is a stability coefficient; is a task integrity constraint; a capacity constraint, a node a maximum physical load, a node a resource elasticity adjustment amount; is a green constraint, is a node green electricity ratio, is a total number of tasks allocated to the node on the node, is a minimum green electricity ratio threshold; An execution feedback module is configured to monitor the actual green electricity utilization rate and dynamically adjust the green reward intensity coefficient to correct the prediction and scheduling deviation.

8. A spatiotemporal prediction-based elastic computing power scheduling method device, comprising a processor and a memory having program instructions stored therein, characterized in that, The processor is configured to execute the method for elastic computing power scheduling based on spatio-temporal prediction according to any one of claims 1-6 when running the program instructions.

Images