An In-Depth Analysis of Advanced Time Series Forecasting Models for the Open RAN

Abstract

Forecasting is instrumental to efficiently manage network resources. In this workshop paper, we make the following contributions. First, we carry out the first assessment of recently proposed advanced forecasting techniques by the AI community, namely DLinear and PatchTST, when applied to the prediction of mobile traffic load and number of users connected to a single Base Station (BS ). We compare these techniques with the well-known Long-Short Term Memory (LSTM) models that are widely adopted in mobile network tasks. Second, we analyze the accuracy tradeoff of these Artificial Intelligence (AI) techniques for single- and multi-step prediction horizons. Third, we profile the operation of all these black-box predictors with an EXplainable Artificial Intelligence ( XAI) lens by using AIChronoLens, a new tool that links legacy XAI explanations with the temporal properties of the input sequences. We find that DLinear excels in single-step horizon predictions while PatchTST and LSTM are more accurate in multi-step horizon predictions. Our XAI study reveals that, unlike PatchTST and LSTM, DLinear focuses its prediction decisions on a few key samples of the input sequences, which ultimately lets it match the ground truth closely.