Event Prediction Large Language Model: Prediction Framework for Highway Abnormal Events Based on Large Language Models

In response to the common issues of fragmented spatiotemporal feature modeling and insufficient interpretability of deep learning models in highway network anomaly event prediction, this paper proposed the event prediction large language model (EP-LLM) framework based on LLMs. The framework employs structured spatiotemporal prompt engineering and parameter-efficient fine-tuning techniques to optimize both prediction accuracy and decision transparency. To address these challenges, we constructed a spatiotemporally aligned multichannel anomaly event dataset to overcome single-source data limitations. We then applied low-rank adaptation to directionally fine-tune the large language model’s decoding layer, enabling multidimensional predictions of congestion levels, accident types, and high-risk intrusions. In addition, a structured prompt template was designed to encode the periodicity, trend, and spatial hotspots of event frequency into interpretable semantic vectors, which was combined with a chain-of-thought reasoning architecture to enhance accuracy. Experimental results demonstrated that EP-LLM substantially outperformed traditional models (long-short term memory LSTM, light gradient-boosting machine, temporal convolutional network) and closely approaches the accuracy of closed-source generative pretrained transformer (GPT)-4o. Specifically, it achieved a 25.4% reduction in mean absolute error (MAE) against LSTM and a 47.3% decrease compared with full-parameter fine-tuning—the latter confirming the substantial advantage of our parameter-efficient design. Ablation studies verified that the synergistic integration of low-rank fine-tuning and structured spatiotemporal prompting collectively accounts for over 47% of the MAE improvement. For multistep forecasting (3–9 steps), the framework reduced error fluctuation by 32.3% while maintaining a stable mean absolute error-root mean square error dispersion coefficient (0.32), demonstrating enhanced robustness. Crucially, EP-LLM delivers these advances with 43% lower computational resource consumption than GPT-4o, achieving an optimal accuracy-efficiency trade-off.