Combined landslide displacement prediction model based on multi-strategy fusion and improved optimization

This study presents a combined prediction model grounded in multi-strategy integrated optimization to tackle the issues of high complexity, non-linearity, and inadequate prediction accuracy in landslide displacement. An improved SFOA incorporating Lévy flight, dynamic exploration adjustment, and stagnation detection enhances global search and convergence. OSFOA is then applied to optimize CEEMDAN, using minimum envelope entropy as the evaluation criterion to reduce hyperparameter subjectivity and decompose cumulative displacement into multi-scale components. The trend term is predicted using a Bayesian-optimized ARIMA model. Periodic and stochastic components are further decomposed by VMD to extract complex frequency features, with Bayesian-optimized SVR models constructed for their prediction. To enhance performance, GRA-MIC is employed to select key influencing factors and optimize input variables for periodic and stochastic terms. This investigation demonstrates that the proposed approach reduces RMSE by 82% and 52% relative to conventional optimization and single decomposition models, respectively. Comparative experiments involving five representative prediction models confirm its superior accuracy and stability over existing methodologies. Additionally, this study introduces an efficient novel framework for complex landslide displacement prediction and early-warning applications.