北京大学学报自然科学版 ›› 2023, Vol. 59 ›› Issue (2): 197-210.DOI: 10.13209/j.0479-8023.2022.106

上一篇    下一篇

基于Transformer的滑坡短期位移预测模型

田原1,2, 庞骁1,2, 赵文祎3,†, 常啸寅1,2, 程楚云1,2, 邹佩4,5, 曹晓澄1,2   

  1. 1. 北京大学遥感与地理信息系统研究所, 北京 100871 2. 空间信息集成与 3S 工程应用北京市重点实验室, 北京 100871 3. 中国地质环境监测院, 100081 4. 北京大学计算机学院, 100871 5. 计算语言学教育部重点实验室, 北京 100871
  • 收稿日期:2022-03-11 修回日期:2022-04-21 出版日期:2023-03-20 发布日期:2023-03-20
  • 通讯作者: 赵文祎, E-mail: 395447712(at)qq.com
  • 基金资助:
    中国地质调查局地质调查项目(DD20211364)和国家重点研发计划(2021YFC3000504-02)资助

A Transformer-Based Model for Short-Term Landslide Displacement Prediction

TIAN Yuan1,2, PANG Xiao1,2, ZHAO Wenyi3,†, CHANG Xiaoyin1,2, CHENG Chuyun1,2, ZOU Pei4,5, CAO Xiaocheng1,2   

  1. 1. Institute of Remote Sensing and Geographical Information Systems, Peking University, Beijing 100871 2. Beijing Key Laboratory of Spatial Information Integration and Its Applications, Beijing 100871 3. China Institute of Geo- Environment Monitoring, Beijing 100081 4. School of Computer Science, Peking University, Beijing 100871 5. Key Lab of Computational Linguistics (MOE), Beijing 100871
  • Received:2022-03-11 Revised:2022-04-21 Online:2023-03-20 Published:2023-03-20
  • Contact: ZHAO Wenyi, E-mail: 395447712(at)qq.com

摘要:

通过将时序卷积网络(TCN)与Transformer解码器进行组合, 提出一种基于Transformer的滑坡短期位移预测模型。将预处理过的位移与降雨序列作为模型的输入, 以时序自回归方式输出未来3日的位移预测结果。实验结果表明, 与支持向量机(SVM)和长短期记忆(LSTM)等传统模型相比, 该模型精度较高, 在快速变形期的预测优势尤为突出。对模型注意力机制的分析结果表明, 模型关注的重点在位移峰值和大降雨附近, 具有较高的可信度。

关键词: 边坡工程, 滑坡位移, 短期预测, Transformer, 注意力机制

Abstract:

A Transformer-based short-term landslide displacement prediction model is proposed by combining temporal convolutional network (TCN) with a Transformer decoder. This model takes the preprocessed displacement and rainfall sequences as input and outputs the displacement predictions for the next three days in a time-series autoregressive manner. The experimental results show that the model achieves higher prediction accuracy than support vector machine (SVM) and long short-term memory (LSTM), and performs particularly well during predicting rapid deformation periods. At the same time, through the analysis of the attention mechanism of the model, it is found that the model focuses on the peak of displacements and heavy rainfalls, indicating that the model is reasonably reliable.

Key words: slope engineering, landslide displacement, short-term prediction, Transformer, attention mechanism