Trend Decomposition Enhanced Large Language Model for Non-Stationary Network Traffic Prediction

Regular network traffic prediction techniques, which generally rely on data normalization, often fail to account for the non-stable characteristics of network traffic, leading to a decline in estimate reliability. A novel approach to this problem is the Trend Decomposition Enhanced Large Language Model for Non-Static Network Traffic Prediction (TD-LLM). Applying trend decomposition as the preparation approach, the proposed strategy first divides raw information into resolved and non-fixed parts. The predetermined portion is standardized and processed, implementing a multi-head cross-attention mechanism, while a better self-attention mechanism improves the non-stationary element. Prompt engineering is used to create specially designed causes that serve as a guideline for the design throughout the forecast process. Finally, the processed data and creates are integrated into the large language model to produce the predicted results. In contrast to foundation models like Time- LLM, LSTM, Informer, and Transformer, the TD-LLM type drastically reduces mean squared error for network traffic prediction across many regions, according to exploratory evaluations, with decreases of 19.73 %, 54.71 %, 76.88 %, and 61.48 %, respectively.