Dynamic prediction of slope displacement using Vmd decomposition with collaborative lssvm-lstm optimization

Area of Science:

• Geotechnical Engineering
• Transportation Infrastructure
• Disaster Prevention

Background:

• China's expanding transportation network increases high-fill and deep-cut subgrade projects.
• Complex geological conditions in these projects elevate landslide risks.
• Accurate slope displacement prediction is crucial for landslide disaster mitigation.

Purpose of the Study:

• To develop a hybrid prediction model (VMLL) for accurate slope displacement forecasting.
• To enhance early warning systems for landslides using small-sample monitoring data.
• To provide a reliable method for assessing slope stability.

Main Methods:

• Proposed the VMD-MPA-LSSVM-LSTM (VMLL) hybrid model.
• Utilized Variational Mode Decomposition (VMD) to separate trend and fluctuation components.
• Employed Least Squares Support Vector Machine (LSSVM) for trend prediction and Long Short-Term Memory (LSTM) for fluctuation prediction.
• Optimized model hyperparameters using the Marine Predators Algorithm (MPA).

Main Results:

• The VMLL model achieved superior prediction accuracy compared to LSSVM, LSTM, and VMD-LSSVM-LSTM.
• Achieved a Mean Absolute Percentage Error (MAPE) of 0.4022% and R² of 94.08% on the Hongtuyao slope dataset.
• Demonstrated high accuracy (MAPE 0.6207%) and robustness across multiple datasets for predicting horizontal displacement and vertical settlement.

Conclusions:

• The VMLL model significantly outperforms traditional methods in slope displacement prediction.
• The hybrid approach offers a robust and accurate framework for landslide early warning.
• This methodology provides a reliable tool for slope stability assessment in transportation infrastructure projects.