A study on the prediction of mountain slope displacement using a hybrid deep learning model

To address the challenges of large prediction errors and limited reliability in conventional modeling approaches, this study proposes a hybrid framework that integrates optimization and deep learning techniques. The method employs an Improved Whale Optimization Algorithm (IWOA) to fine-tune parameters for GNSS data fitting, ensuring accurate signal feature extraction. These parameters are then fed into a Long Short-Term Memory (LSTM) network to model spatiotemporal dependencies through deep temporal pattern learning. Finally, a Gradient Boosted Decision Trees (GBDT) module is used to correct residual errors, particularly for predictions with large deviations, thereby improving overall accuracy and robustness. Unlike conventional models, this hybrid framework effectively mitigates large errors and improves reliability by leveraging a multi-stage approach. Experimental results confirm that the IWOA-LSTM-GBDT framework significantly outperforms traditional models. On the long-term prediction task at station JC03, it achieves a 37.9% reduction in Root Mean Square Error (RMSE), a 32.4% decrease in Mean Absolute Error (MAE), and a 4.6% increase in the coefficient of determination (R2) compared to the baseline LSTM. Compared to the IWOA-LSTM variant without residual correction, the complete framework further reduces RMSE by 7.2%, MAE by 2.3%, and increases R2 by 0.5%. However, the framework may require significant computational resources, and its performance may be sensitive to the quality of input data, particularly for stations with limited measurements.