Research On Intelligent Mine Network Traffic Prediction And Optimization Algorithms

The analysis,prediction,and optimization of network traffic are closely related.By exploring the characteristics of historical traffic time series data,identifying traffic patterns and distributions in the network,and making predictions,more accurate mathematical models or heuristic search spaces can be established to optimize network efficiency and quality.In recent years,intelligent coal mines have been rapidly developing,and new coal mine networks or information architectures,such as network function service-oriented networking,hierarchical microservice collaboration,and cloud-edge joint network resource management,have been frequently mentioned.which aim to improve the shortcomings of existing network architectures in terms of latency,reliability,and business collaboration flexibility,to meet the high-quality ondemand collaboration needs of coal mine industrial Io T services.In this regard,the application of network function virtualization and software-defined networking can conveniently achieve global network measurement,network configuration,and network service orchestration.However,there is still a lack of intelligent demand cognition tools and network service optimization methods between network status awareness and configuration actions.The current dynamic of coal mine network traffic contains deep features of the entire production process and life cycle information interaction,such as time dependence in different periods,spatial correlation,terminal mobility,business dependence,and other latent upper-layer application interaction semantics.Effectively capturing these features and dynamically deploying the most reasonable services on limited resources has become a key scientific issue in the construction of intelligent coal mines.Therefore,this dissertation studies the relevant predictive models for complex dynamic traffic in the context of intelligent mining.Building upon precise identification of network features,further research is conducted on the deployment of network virtual services and the performance assurance in the dynamic cloud-edge scenarios of mining.The main contributions of the dissertation are as follows:1)A network traffic prediction model that combines network time and space features was proposed to address the overlapping of temporal and spatial variations in network traffic.This model introduces graph convolutional networks to extract spatial dependencies of network node traffic on the link topology,and utilizes gated recurrent units to further capture the temporal evolution of traffic spatial latent features.It effectively characterizes the spatiotemporal correlation features arising from terminal mobility and random interactions.Experimental results demonstrate that compared to other prediction methods based solely on temporal or spatial features,this model exhibits superior performance in predicting aggregated network traffic sequences.2)A method for online fast adaptive feature meta-learning for short-term traffic prediction is proposed to address the issue of insufficient characteristics of user shortterm arrival traffic sequences.This method does not depend on the assumption of the training and testing data having the same distribution as traditional deep learning traffic prediction algorithms.Instead,during the meta-training phase,it alternately learns the intra-series features and inter-series process differences to optimize the global initialization parameters of the model.During the meta-testing phase,it dynamically matches the best sub-predictor based on error feedback,achieving rapid adaptation of new features and enhancing the model’s generalization ability for predicting random non-stationary traffic.3)In response to the demand for fine-grained resource management of virtualized network services for key business in coal mine networks,a general network performance prediction and optimization method adaptable to different network topologies,routing plans,and traffic matrix combinations is proposed.This method is based on an elastic network management framework,abstracting the traffic perception,prediction,and optimization decision-making processes.Firstly,a graph neural network based deep learning model is constructed to capture complex spatiotemporal hidden relationships between network topology,routing,and business features,which enhances the prediction accuracy of network performance indicators.Secondly,the performance indicators prediction model is embedded in the deep reinforcement learning model for network function load optimization,achieving adaptive optimization decisions through precisely network environment awareness.Experimental results demonstrate a significant performance improvement compared to existing methods,with stronger generalization ability across different topological structures.4)A dynamic multi-task slicing services function chain proactive deployment method was proposed to address the challenges of diverse services,spatial-temporal variations in user functional demands,uncertain impact of time-varying traffic on heterogeneous VNF resource costs,and difficulties in edge resource adaptation in the context of mining industrial Io T network slicing.Firstly,to address the issue of user business demands being influenced by multiple factors,a multi-graph convolutional gated recurrent unit is introduced to extract the hidden spatial-temporal features of user business demands by integrating heterogeneous spatial information and capturing the disturbance of external events in mining production on business demands using an attention mechanism.This approach better captures the dynamic changes in user demands,providing accurate spatial distribution and load intensity information for slicing deployment.Additionally,to capture the heterogeneity of network functions in resource demands and the non-linear characteristics of resource consumption under different loads,a meta-learning-based mapping method for user traffic load and virtualized network function service resource demands is proposed to improve the accuracy of predicting heterogeneous VNF resource consumption.Finally,aiming to maximize network resource utilization and user access satisfaction,a mining industrial Io T cloud-edge slicing service chain dynamic deployment problem model is constructed,and a parallel multi-task reinforcement learning slicing service deployment algorithm that combines traffic and resource prediction is proposed to effectively alleviate the action space exploration problem in the slicing process for function chain deployment and resource allocation.Simulation results demonstrate that the proposed method can maintain better user acceptance rates,resource utilization,and convergence performance in various business scenarios,outperforming other comparative methods.The dissertation contains 47 figures,18 tables and 171 references.