Abstract: The prediction of seismic responses in structural engineering is a critical component of building assessment, particularly in performance-based seismic engineering. This paper focuses on scenarios with limited data samples in seismic response prediction and proposes a method termed LLM-PaP (large language model with prompt-as-prefix). This method leverages the general analytical capabilities of large language models (LLMs) for time-series data to address performance deficiencies typically encountered by conventional models under small sample conditions. The model integrates the “PaP (prompt-as-prefix)” concept, enhancing the understanding of input sequences by incorporating natural language task instructions and statistical information of seismic input sequence data to guide the reasoning and prediction process. Experimental validation conducted on two datasets demonstrates the effectiveness of the proposed method. The results indicate that LLM-PaP significantly outperforms advanced prediction methods based on MLP, frequency domain analysis, and Transformer models in terms of predictive performance on the datasets. Additional experiments on generalization reveal the superior adaptability of LLM-PaP across datasets. LLM-PaP presents an innovative solution for seismic response prediction tasks, offering new insights and methods for the interdisciplinary research at the large models and seismic response prediction in the future.

Key words: seismic response prediction, large language model (LLM), prompt learning, time series

摘要： 建筑结构的地震响应预测是基于性能的地震工程中建筑评估的重要组成部分。聚焦于地震响应预测中的少样本场景，提出了一种融合了大语言模型（large language model，LLM）和提示学习的震动响应预测方法LLM-PaP。该方法将LLM对时序数据的通用分析能力迁移到地震响应预测中，以克服一般模型在小样本条件下的性能缺陷。在模型中引入“PaP（prompt-as-prefix）”思想，为时间序列增加自然语言任务指令和地震输入序列数据的统计信息，以增强模型对输入序列的理解能力并引导推理预测过程。在两个数据算例的实验中验证了所提方法的有效性。结果表明：LLM-PaP在数据集上的预测性能显著优于基于MLP、频域和Transformer等先进预测方法。进一步的泛化性实验结果揭示了LLM-PaP在跨数据集适应上的卓越性能。LLM-PaP为地震响应预测任务提供了一种创新性的解决方案，为未来大模型与震动响应预测领域的交叉性研究提供了新的思路和方法。

关键词: 地震响应预测, 大语言模型（LLM）, 提示学习, 时间序列