Network KPI Prediction in Mobile Networks Using CNN-LSTM Time Series Model

Accurate prediction of network key performance indicators (KPIs), such as throughput and latency, is important for maintaining Quality of Service (QoS) and enabling intelligent, proactive optimization in mobile applications to support reliable and ubiquitous connectivity. One application is the monitoring of livestock transport, a key focus of the Horizon Europe project<br/>COMMECT, which aims to develop a smart multi-connectivity solution to ensure seamless connectivity throughout the transport route. Although beneficial, predicting network KPIs based on<br/>radio KPIs is particularly challenging in dynamic environments due to the intricate nature of their relationship. In our previous study, this weak relationship led to limited performance in applying a supervised Random Forest (RF) model. Aiming to improve the prediction of network KPIs, we investigate the convolutional neural network-long short-term memory (CNNLSTM) time-series model, which combines convolutional layers for spatial feature extraction with LSTM layers to learn temporal dependencies. Unlike the supervised RF model, the current approach predicts network KPIs based on past information about the network KPI itself, bypassing the need for a mapping between radio and network KPIs. In our comparison, the CNN-LSTM time series and RF models have been trained and tested based on measured 5G throughput and latency data collected in northern Denmark. The CNN-LSTM model outperforms the RF model, achieving an R2 score improvement of 26.17% for latency and 15.81% for throughput, highlighting its ability to adapt to varying network conditions in a realistic setting.