用深度学习预测盐岩蠕变，提升地下储能设施安全评估精度
Salt-Rock Creep Deformation Forecasting Using Deep Neural Networks and Analytical Models for Subsurface Energy Storage Applications
- 结合深度神经网络与解析模型，捕捉盐岩长期蠕变规律
- N-BEATS和TCN模型比传统方法准确率提升15%-20%
- 适合地下储氢、核废料储存等工程的长期稳定性评估
本研究深入分析时间序列预测方法，以预测在不同围压条件下盐岩的时变变形（即蠕变）趋势。蠕变评估对设计和运营核废料、氢能或放射性物质的地下储存设施至关重要。采用多级三轴蠕变数据，轴向应变数据以5–10秒间隔记录，围压范围为5–35 MPa，持续时间5.8–21天。初步分析显示轴向应变与温度间无显著季节性或因果关系。增广迪基-富勒（ADF）检验表明数据平稳（p<0.05），小波相干分析揭示重复趋势。对比了多种深度神经网络模型（N-BEATS、TCN、RNN、Transformer）与统计基准模型，使用RMSE、MAE、MAPE、SMAPE评估性能。结果表明，N-BEATS和TCN在不同应力水平下表现最优，较传统解析模型提升15%–20%精度，有效捕捉复杂时序依赖与模式。
原文摘要 · Abstract (English)
This study provides an in-depth analysis of time series forecasting methods to predict the time-dependent deformation trend (also known as creep) of salt rock under varying confining pressure conditions. Creep deformation assessment is essential for designing and operating underground storage facilities for nuclear waste, hydrogen energy, or radioactive materials. Salt rocks, known for their mechanical properties like low porosity, low permeability, high ductility, and exceptional creep and self-healing capacities, were examined using multi-stage triaxial (MSTL) creep data. After resampling, axial strain datasets were recorded at 5--10 second intervals under confining pressure levels ranging from 5 to 35 MPa over 5.8--21 days. Initial analyses, including Seasonal-Trend Decomposition (STL) and Granger causality tests, revealed minimal seasonality and causality between axial strain and temperature data. Further statistical tests, such as the Augmented Dickey-Fuller (ADF) test, confirmed the stationarity of the data with p-values less than 0.05, and wavelet coherence plot (WCP) analysis indicated repeating trends. A suite of deep neural network (DNN) models (Neural Basis Expansion Analysis for Time Series (N-BEATS), Temporal Convolutional Networks (TCN), Recurrent Neural Networks (RNN), and Transformers (TF)) was utilized and compared against statistical baseline models. Predictive performance was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Symmetric Mean Absolute Percentage Error (SMAPE). Results demonstrated that N-BEATS and TCN models outperformed others across various stress levels, respectively. DNN models, particularly N-BEATS and TCN, showed a 15--20\% improvement in accuracy over traditional analytical models, effectively capturing complex temporal dependencies and patterns.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。