A Performance Prediction Method Based on Multi-task Spatio-Temporal Convolution Network for SDN Heterogeneous Network

As a large number of terminal devices are deployed in heterogeneous network based on software-defined networking (SDN), the complexity of topology and dynamic change of performance parameters (e.g., latency, load, jitter, etc.) increase importantly, leading to the difficulties in network performance prediction. To address these issues, a performance prediction method for SDN-enabled heterogeneous network is proposed. First, according to the acquired real-time measurement data from SDN-enabled heterogeneous network, a graph neural networks (GNN) feature extraction model based on temporal gated networks is constructed to obtain feature vectors for network performance prediction. Second, considering the coupling relationship of network performance parameters, a multi-task learning method based on multi-expert gated networks is proposed to accurately achieve the performance prediction of network. Finally, experimental results confirm the feasibility and effectiveness of the proposed prediction method.