Temperature Prediction of Seasonal Frozen Subgrades Based on CEEMDAN-LSTM Hybrid Model

Area of Science:

• Geotechnical Engineering
• Environmental Science
• Artificial Intelligence

Background:

• Accurate subgrade temperature prediction is vital for understanding thermal states in seasonally frozen regions.
• Subgrade temperature time series exhibit nonlinearity and non-stationarity, challenging traditional neural networks.
• Hybrid models offer improved forecasting capabilities for complex time series data.

Purpose of the Study:

• To enhance the accuracy of subgrade temperature prediction in seasonally frozen regions.
• To evaluate the effectiveness of hybrid models in capturing nonlinear and non-stationary temperature data.
• To develop and validate a novel CEEMDAN-LSTM model for subgrade thermal state analysis.

Main Methods:

• Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD), and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) were applied to temperature time series.
• A hybrid dataset was constructed incorporating volumetric water content and frost heave time series.
• The CEEMDAN-LSTM model was developed and trained using the hybrid dataset for temperature prediction.

Main Results:

• The CEEMDAN-LSTM model demonstrated superior prediction performance compared to other decomposed LSTM models.
• The inclusion of a hybrid dataset significantly improved the predictive ability of the models.
• The CEEMDAN-LSTM model effectively addressed the nonlinearity and non-stationarity inherent in subgrade temperature time series data.

Conclusions:

• The CEEMDAN-LSTM model offers a promising approach for accurate subgrade temperature forecasting in seasonally frozen environments.
• Hybrid datasets enhance the capability of deep learning models to handle complex environmental time series.
• This study contributes to a better understanding and prediction of thermal states in seasonally frozen subgrades.