TLS-Net: A Hybrid Time Series Prediction Model Combining TCN and LSTM for Ship-Satellite Network Traffic

Accurate prediction of ocean-going ship satellite network traffic (OSSNT) can optimize communication resource allocation and improve communication efficiency. But ordinary statistical learning models cannot capture the non-linear relationships between data well. Standalone single deep learning prediction models cannot extract short-term features and capture long-term dependence in time series data simultaneously. In addition, traffic forecasting studies for ship satellite networks have not been investigated due to the unavailability of data. In this paper, a real-world situation-based ship-satellite network traffic dataset is constructed. Furthermore, a hybrid time series prediction neural network model (TLS-Net), which is built on Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM), is proposed to forecast OSSNT. This model learns short-term local feature in the time series from TCN, and then captures the long-term dependence of relevant data by LSTM, which extracts multi-source temporal feature and high and low-frequency information. Considering the high noise and non-linear in the raw data, we use Savitzky-Golay filters to eliminate potential noise. Experimental results demonstrate that the proposed model predicts network traffic more accurately than other vanilla prediction models.