Abstract:

To address the challenges that static prediction models face in accurately forecasting dynamic landslide trends and the high computational costs associated with dynamic models—which hinder real-time prediction—this study proposes a novel real-time model for landslide displacement prediction. The model integrated Particle Swarm Optimization (PSO), a Gated Recurrent Unit (GRU) network, Online Learning (OL), and dynamic Model Retraining (MRT). First, a static landslide prediction model was established by integrating PSO and GRU. Second, an OL strategy was incorporated into the static model, enabling dynamic updates and real-time predictions as new monitoring data were acquired. Then, small-batch MRT was performed based on prediction accuracy evaluation to predict landslide trends dynamically and in real time. Finally, a comparative analysis of several related models was conducted using the Wangeryan landslide in Sichuan Province as a case study. The results indicate that the OL and MRT methods significantly improve prediction accuracy. Specifically, the PSO-GRU-OL-MRT model achieved Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and R² values of 0.795, 3.53, 1.40, and 0.954, respectively, with an average prediction time of 25.0 seconds per instance, demonstrating the highest prediction accuracy. In comparison, the GRU-OL-MRT model yielded values of 1.73, 7.82, 2.54, and 0.917 for the same four metrics, with an average prediction time of 0.496 seconds per instance, significantly reducing computational costs while maintaining relatively high prediction accuracy.

Key words: particle swarm optimization (PSO), gated recurrent unit (GRU), online learning (OL), dynamic model re-train (MRT), landslide displacement, real-time prediction