基于LSTM与Transformer的地面沉降智能预测方法研究——以上海市为例

Intelligent prediction of land subsidence based on LSTM and Transformer:A case study of Shanghai

受地面沉降严重威胁到生命财产安全的人口已达19%,开展地面沉降模拟预测对防灾减灾具有非常重要的现实意义。针对现有地面沉降预测在模型参数难以获取、单一深度学习方法在预测精度低等方面的局限性,本文提出了集成大模型核心技术的地面沉降预测方法。首先,从地面沉降模拟预测的顶层设计,提出了基于深度学习的地面沉降预测包括算力层、数据层、模型层、评估层与应用层的总体架构;其次,基于LSTM与Transformer提出了地面沉降预测的实用方法;最后,利用上海的地面沉降数据进行了实验研究。结果表明:深度学习技术可以在地面沉降模拟预测中取得较好的结果,多模型法对地面沉降变化不大、回弹、变化较大均可进行预测,iTransformer模型对地面沉降变化较小的情况预测效果较好;在微量地面沉降时代,利用大模型的核心技术Transformer可以取得较高的精度。

Land subsidence poses a serious threat to the lives and property of 19%of the world’s population.It is of great practical significance to carry out land subsidence simulation and prediction for disaster prevention and mitigation.With the advancement of large language model technology,it becomes imperative to explore the application in the field of land subsidence.Addressing the limitations of the existing land subsidence prediction models,which struggle with parameter acquisition and single-method accuracy,this paper proposes a land subsidence prediction method integrating the core technology of large language models.Commencing with the top-level design of land subsidence simulation prediction,this paper proposes the overall architecture based on depth learning.This architecture encompasses the computing power layer,data layer,model layer,evaluation layer,and application layer.The computing power layer involves hardware and distributed deep learning frameworks,integrating GPUs with deep learning frameworks for efficient parallel computing.In the data layer,land subsidence data is mainly obtained through leveling,bedrock benchmark,stratification benchmark,automatic monitoring,GNSS,and InSAR technologies.The model layer utilizes various deep learning models such as RNN,LSTM,Transformer,TimeGPT,and hybrid models to analyze and predict land subsidence monitoring data.In the evaluation layer,multiple deep learning models are assessed against measured data to obtain the optimal method,ultimately selecting the best model for prediction.At the application layer,prediction results can provide decision-making support for urban safety,major infrastructure structure safety,land subsidence monitoring and early warning,and disaster prevention and mitigation.Secondly,the paper proposes a practical approach to land subsidence prediction based on the core technology of large language models,specifically LSTM and Transformer.Finally,Shanghai’s land subsidence were used for experimental research.Based on the Shanghai Geological Environment Bulletin and our institute’s“Geological Environment Monitoring Information Management Platform”,18 monitoring stations were selected from key control areas of land subsidence,sub-key control areas,and general control areas.These stations have over 10 years of leveling data with good data integrity.The data serves as sources for simulation and prediction analysis.The land subsidence prevention and control area’s scope map is derived from the geological environment bulletin.Monthly on-site measurements conducted by our organization ensure the accuracy of leveling data,meeting survey specifications and ensuring reliable data.With the exception of A2,the land subsidence prediction methods proposed in this paper have achieved good results.More importantly,the iTransformer model has demonstrated effective predictions for monitoring stations with minimal changes in sedimentation.The results show that deep learning technology can achieve good results in land subsidence simulation and prediction.The multi-model method proves capable of predicting scenarios with minimal change,rebound,and substantial change in land subsidence.The iTransformer model has a good prediction effect on the situation with small changes in land subsidence.In the era of micro land subsidence,high accuracy can be achieved through the utilization of Transformer,the core technology of large language models.