Hybrid model integrating multi-stage machine learning optimization for accurate landslide displacement prediction

Accurate prediction of landslide displacement faces severe challenges due to its high nonlinearity and the complexity of multi-source influencing factors, which limits the prediction accuracy and robustness of existing models. To address this issue, this study proposes a systematic, Multi-stage Optimized Hybrid Prediction Model (MOHPM). The core of this model lies in constructing a collaborative optimization framework aimed at systematically tackling the computational overhead, sensitivity to data anomalies, and parameter optimization difficulties encountered by complex hybrid models in practical applications. The framework first builds a base model that integrates a Temporal Convolutional Network (TCN), Principal Component Analysis (PCA), and Support Vector Regression (SVR). On this basis, to comprehensively improve overall performance, the study innovatively incorporates a set of targeted optimization strategies: a pruning algorithm (PA) is used to reduce computational complexity; the Local Outlier Factor (LOF) is applied to identify and handle data anomalies to enhance robustness; and Simulated Annealing (SA) is employed for global hyperparameter optimization. Experiments on the public dataset of the Three Gorges Reservoir demonstrate that the final MOHPM model exhibits outstanding performance. Compared with the benchmark single SVR model, the Mean Absolute Error (MAE) decreases by over 80%. The prediction coefficient of determination (R2) on the test set reaches 0.996, with a missed detection rate and false alarm rate as low as 0.87% and 6.81%. Crucially, in robustness tests with injected random noise, the average prediction accuracy decreases by only 2.4% points, highlighting its high stability under complex conditions. The proposed framework provides a new methodology for precise landslide early warning, with significant theoretical innovation and practical application potential.