CN-121979692-A - Intelligent resource allocation method and system for railway cloud platform

CN121979692ACN 121979692 ACN121979692 ACN 121979692ACN-121979692-A

Abstract

The invention provides an intelligent resource allocation method and system for a railway cloud platform, the method comprises the steps of preprocessing resource index data to obtain a resource feature sequence, establishing a three-layer LSTM prediction model, defining a current time resource feature sequence according to the resource feature sequence, carrying out forward calculation by combining the three-layer LSTM prediction model to obtain a future time resource feature sequence, defining a historical time resource feature sequence according to the resource feature sequence, establishing a comprehensive system state vector based on the historical time resource feature sequence, the current time resource feature sequence and the future time resource feature sequence, respectively establishing a double DQN decision model and an experience data quadruple, carrying out optimization training on the double DQN decision model through the experience data quadruple, carrying out forward calculation based on the double DQN decision model, outputting an optimal resource allocation decision, and carrying out decision checking on the optimal resource allocation decision. The invention can improve the utilization rate of the whole resources of the railway cloud platform.

Inventors

ZHAO DOU
WANG ZHIBIN
ZHOU FENG
HAO JUN
ZHENG YUJIE
WANG SHUJIE
Nan Bingshen
Jia Peiran
SHAN XINYI
ZHAO LIUJUN
YU XINGJIAN
PENG LIANGYONG
LI JINGQIN
WEI YICHEN
Ran jie
GAO BIN
LIU WEIZHUANG

Assignees

中国铁路设计集团有限公司

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (9)

1. An intelligent resource allocation method for a railway cloud platform is characterized by comprising the following steps: acquiring resource index data, and preprocessing the resource index data to obtain a resource characteristic sequence; Establishing a three-layer LSTM prediction model, and carrying out optimization training on the three-layer LSTM prediction model by utilizing a resource feature sequence; Defining a current time resource feature sequence according to the resource feature sequence, and performing forward calculation by combining the current time resource feature sequence with a three-layer LSTM prediction model to obtain a future time resource feature sequence; defining a historical moment resource feature sequence according to the resource feature sequence, and establishing a comprehensive system state vector based on the historical moment resource feature sequence, the current moment resource feature sequence and the future moment resource feature sequence; Establishing a double DQN decision model consisting of an online Q network and a target Q network, defining a t moment comprehensive system state vector according to the comprehensive system state vector, acquiring t+1 moment resource index data, establishing an experience data quadruple based on the t moment comprehensive system state vector and the t+1 moment resource index data, and carrying out optimization training on the double DQN decision model through the experience data quadruple; inputting the comprehensive system state vector at the time t into a double DQN decision model for forward calculation, and outputting an optimal resource allocation decision; And carrying out decision checking on the optimal resource allocation decision.
2. The intelligent resource allocation method for a railway cloud platform according to claim 1, wherein the optimizing training of the three-layer LSTM prediction model by using the resource feature sequence comprises: Defining a resource feature sequence as Inputting the resource feature sequence into a three-layer LSTM prediction model, outputting a sample resource demand prediction result with 6 future time steps, comparing the sample resource demand prediction result with a sample resource demand true value, and calculating a prediction error; And quantizing the prediction error through a first mean square error loss function, optimizing network parameters of the three-layer LSTM prediction model through a back propagation algorithm, and repeating the iterative updating until the first mean square error loss function is converged, wherein when the first mean square error loss function is converged, the three-layer LSTM prediction model optimization training is completed.
3. The intelligent resource allocation method for the railway cloud platform according to claim 2, wherein a Dropout mechanism with a drop rate of 0.1 is introduced in the process of optimizing and training the three-layer LSTM prediction model so as to improve the generalization capability of the three-layer LSTM prediction model.
4. The intelligent resource allocation method for the railway cloud platform according to claim 1, wherein defining a current time resource feature sequence according to the resource feature sequence, performing forward computation by combining the current time resource feature sequence with a three-layer LSTM prediction model, and obtaining a future time resource feature sequence includes: defining the current time resource feature sequence as the resource feature sequence according to the resource feature sequence The current time resource feature sequence Inputting the result into a three-layer LSTM prediction model for forward calculation, and outputting a predicted result of the current time resource demand with 6 future time steps Selecting the last time step As a current time resource feature sequence And (3) a predicted value corresponding to 1 hour later, wherein the predicted value is the future time resource characteristic sequence.
5. The intelligent resource allocation method for a railway cloud platform according to claim 1, wherein defining a historical time resource feature sequence, a current time resource feature sequence, and a future time resource feature sequence according to the resource feature sequence to establish a comprehensive system state vector comprises: Defining a historical time resource feature sequence according to the resource feature sequence, and respectively setting the historical time resource feature sequence as The characteristic sequence of the current time resource is as follows The future time resource feature sequence is Splicing the historical time resource feature sequence, the current time resource feature sequence and the future time resource feature sequence along feature dimensions to establish and obtain a comprehensive system state vector with the dimension number of 3m, wherein the expression is as follows: 。
6. The intelligent resource allocation method for the railway cloud platform according to claim 1, wherein defining a t moment comprehensive system state vector according to the comprehensive system state vector and obtaining t+1 moment resource index data, establishing an experience data quadruple based on the t moment comprehensive system state vector and the t+1 moment resource index data, and performing optimization training on the double DQN decision model through the experience data quadruple comprises: defining the comprehensive system state vector at the time t as according to the comprehensive system state vector Inputting the state vector of the t-moment comprehensive system into a double DQN decision model to obtain t-moment resource scheduling action Bonus function at time t ; Acquiring t+1 time resource index data, preprocessing the t+1 time resource index data to obtain a t+1 time resource feature sequence, performing forward calculation by combining the t+1 time resource feature sequence with a three-layer LSTM prediction model to obtain a predicted t+2 time resource feature sequence, and establishing a t+1 time comprehensive system state vector based on the t time resource feature sequence, the t+1 time resource feature sequence and the predicted t+2 time resource feature sequence ; Comprehensive system state vector based on t moment Comprehensive system state vector at time t+1 Resource scheduling actions at time t Instant rewards at time t Building a quadruple of empirical data Storing the experience data quadruple into an experience playback pool, randomly selecting part of experience data quadruple data from the experience playback pool as training data, and carrying out iterative updating on an online Q network in the double DQN decision model based on the training data; Integrating the system state vector at the time t Input in online Q network, output t time resource scheduling action Integrating the t+1 moment into a system state vector Input in online Q network, output t+1 time resource scheduling action A second predicted Q value; enabling an online Q network to select t+1 moment resource scheduling action with the maximum second predicted Q value Scheduling actions by the target Q network on the t+1 time resource with the maximum second predicted Q value Performing value evaluation and rewarding function at time t Adding to obtain a target Q value; And minimizing the error between the first predicted Q value and the target Q value, updating the network parameters of the online Q network through a back propagation algorithm, synchronizing the network parameters of the online Q network to the target Q network according to a preset period, and updating the network parameters of the target Q network.
7. The intelligent resource allocation method for a railway cloud platform as claimed in claim 6, wherein the online Q network is defined as The target Q network is The expression for calculating the target Q value using the bellman update formula is: ; In the formula, Is a discount factor for balancing immediate returns with long-term returns; And Network parameters of the online Q network and the target Q network, respectively.
8. The intelligent resource allocation method for the railway cloud platform according to claim 1, wherein the step of inputting the t-moment comprehensive system state vector into the double DQN decision model for forward calculation and outputting the optimal resource allocation decision comprises the step of inputting the t-moment comprehensive system state vector into the double DQN decision model for forward calculation to obtain the resource scheduling actions at each t moment And selecting the t-time resource scheduling action with the largest first predicted Q value And taking the optimal resource allocation decision as an optimal resource allocation decision, and outputting the optimal resource allocation decision.
9. An intelligent resource allocation system for a railway cloud platform, comprising: The data processing unit is used for acquiring the resource index data, and preprocessing the resource index data to obtain a resource characteristic sequence; The three-layer LSTM prediction model optimization training unit is used for establishing a three-layer LSTM prediction model and carrying out optimization training on the three-layer LSTM prediction model by utilizing a resource feature sequence; The predicted resource feature sequence acquisition unit is used for defining a current time resource feature sequence according to the resource feature sequence, and performing forward calculation by combining the current time resource feature sequence with a three-layer LSTM prediction model to obtain a future time resource feature sequence; the comprehensive system state vector acquisition unit is used for defining a historical time resource feature sequence according to the resource feature sequence, and establishing a comprehensive system state vector based on the historical time resource feature sequence, the current time resource feature sequence and the future time resource feature sequence; The double DQN decision model optimization training unit is used for establishing a double DQN decision model consisting of an online Q network and a target Q network, defining a t moment comprehensive system state vector according to the comprehensive system state vector, acquiring t+1 moment resource index data, establishing an experience data quadruple based on the t moment comprehensive system state vector and the t+1 moment resource index data, and performing optimization training on the double DQN decision model through the experience data quadruple; the optimal resource allocation decision acquisition unit is used for inputting the comprehensive system state vector at the moment t into the double DQN decision model to perform forward calculation and outputting an optimal resource allocation decision; And the decision checking unit is used for checking the optimal resource allocation decision.

Description

Intelligent resource allocation method and system for railway cloud platform Technical Field The invention relates to the technical field of cloud computing resource management and intelligent scheduling, in particular to an intelligent resource allocation method and system for a railway cloud platform. Background With the continuous development of cloud computing technology and virtualization technology, a cloud platform becomes an important infrastructure for supporting the operation of a large-scale service system, and can provide resource support adapted as required for various railway services, so that the resource utilization efficiency is improved, and the system operation cost is reduced. At present, a railway cloud platform is used as a core information infrastructure to intensively bear a plurality of key business systems such as transportation scheduling, passenger ticket business, passenger service, video monitoring, equipment overhaul and the like. The business has the characteristics of high concurrency degree, strong real-time requirement, obvious load fluctuation and the like in the running process, wherein part of the business has strict requirements on the system response time and the continuous running capability, the safety and the quality of railway transportation are directly related, and meanwhile, the railway business running has obvious time regularity and burstiness characteristics, such as the scenes of passenger transport peak, holiday concentrated traveling, emergency disposal and the like, and can greatly influence cloud platform resources in a short time. Under the complex service scene, the cloud platform not only needs to ensure the stable operation of the key service under the high-load condition, but also needs to realize reasonable coordination and efficient scheduling among multiple services under the condition of limited resources. Disclosure of Invention In view of the above, the invention aims to provide an intelligent resource allocation method and system for a railway cloud platform, which can improve the overall resource utilization rate of the railway cloud platform. In a first aspect, an embodiment of the present invention provides an intelligent resource allocation method for a railway cloud platform, including acquiring resource index data, and preprocessing the resource index data to obtain a resource feature sequence; Establishing a three-layer LSTM prediction model, and carrying out optimization training on the three-layer LSTM prediction model by utilizing a resource feature sequence; Defining a current time resource feature sequence according to the resource feature sequence, and performing forward calculation by combining the current time resource feature sequence with a three-layer LSTM prediction model to obtain a future time resource feature sequence; defining a historical moment resource feature sequence according to the resource feature sequence, and establishing a comprehensive system state vector based on the historical moment resource feature sequence, the current moment resource feature sequence and the future moment resource feature sequence; Establishing a double DQN decision model consisting of an online Q network and a target Q network, defining a t moment comprehensive system state vector according to the comprehensive system state vector, acquiring t+1 moment resource index data, establishing an experience data quadruple based on the t moment comprehensive system state vector and the t+1 moment resource index data, and optimally training the double DQN decision model by inputting the t moment comprehensive system state vector into the experience data quadruple; The double DQN decision model carries out forward calculation and outputs an optimal resource allocation decision; And carrying out decision checking on the optimal resource allocation decision. Further, performing optimization training on the three-layer LSTM prediction model by using the resource feature sequence comprises the following steps: Defining a resource feature sequence as Inputting the resource feature sequence into a three-layer LSTM prediction model, outputting a sample resource demand prediction result with 6 future time steps, comparing the sample resource demand prediction result with a sample resource demand true value, and calculating a prediction error; And quantizing the prediction error through a first mean square error loss function, optimizing network parameters of the three-layer LSTM prediction model through a back propagation algorithm, and repeating the iterative updating until the first mean square error loss function is converged, wherein when the first mean square error loss function is converged, the three-layer LSTM prediction model optimization training is completed. Furthermore, in the process of carrying out optimization training on the three-layer LSTM prediction model, a Dropout mechanism with the discarding rate of 0.1 is introduced to improve