Large Language Models (LLMs) for Network Traffic Prediction: A Trend-Aware Hybrid Framework

Yuzhou Chen, Kwok-Yan Lam, Feng Li

Abstract: The explosive growth and increasing complexity of modern 5G/6G networks, driven by Internet of Things (IoT), industrial automation, and real-time multimedia streaming, demand forecasting methods that address nonstationarity, abrupt shifts, and incomplete observations. Nonstationarity involves changing statistical properties, abrupt shifts stem from events or outages, and data gaps can impair model accuracy. Traditional statistical models and deep sequence learners partially handle these challenges but often leave systematic residuals, which are structured errors from unmodeled scenarios and overlook high-level contextual cues such as external events or semantic patterns. To overcome these limitations, we propose a hybrid forecasting framework combining a convolutional neural network-long short-term memory (CNN-LSTM) trend predictor for capturing local fluctuations and long-range dependencies, an extreme gradient boosting (XGBoost) residual corrector to refine forecast errors, and a low-rank adaptation (LoRA)-fine-tuned large language model (LLM) that generates semantic labels of trend direction, anomaly type, and volatility regime to enrich residual learning. Experimental evaluation on real-world cellular traffic data shows up to 15% reduction in root-mean-square error (RMSE) and 10% reduction in mean absolute percentage error (MAPE) compared to state-of-the-art hybrid baselines, with substantially improved resilience to noise, missing data, and abrupt traffic surges. Our contributions include a parameter-efficient prompt-based LoRA fine-tuning pipeline for adapting LLMs to time-series forecasting, a context-aware residual learning architecture fusing numerical and linguistic features, and comprehensive empirical validation demonstrating superior accuracy and robustness in dynamic network environments.

External IDs:dblp:journals/iotj/ChenLL26