Multi-Source Heterogeneous Data-Driven Intelligent Prediction for Landslide Dam Longevity

Abstract

Accurate prediction of the longevity of Ldam, as temporary or permanent hydraulic structures, is crucial for ensuring downstream safety of lives and properties. This study proposes an intelligent forecasting model to accurately predict the longevity of Ldam. Firstly, a database containing a large amount of Ldam data is collected and organized, with key factors selected as indicators for the prediction system. Statistical indicators of the database are calculated, and in-depth analysis is conducted using correlation heatmaps and violin plots. Secondly, an intelligent forecasting model is constructed based on an improved intelligent optimization algorithm and ensemble learning. The model consists of multiple base learners (MLP, SVR, CatBoost) and a meta-learner (LightGBM). To effectively improve model performance, an improved intelligent optimization algorithm called IGTO is proposed to optimize the hyperparameters of the meta-learner within the stacked ensemble learning framework. In the analysis of the model's prediction results, prediction plots and regression plots are provided, and a series of evaluation metrics (R2, Adj-R2, RMSE, MAE, MAPE, VAF) are calculated. The results demonstrate that the IGTO stacked model exhibits high accuracy and reliability in predicting the longevity of Ldam, with computed values of R2 = 0.98, Adj-R2 = 0.98, RMSE = 93.98, MAE = 48.59, MAPE = 0.46, VAF = 98.29, showing good agreement with actual observed values. Furthermore, the model outperforms other prediction models and previous empirical formulas, validating the effectiveness and practicality of the IGTO improved optimization algorithm and ensemble learning framework. Additionally, the SHAP method is employed to assess the importance and impact of each input parameter on the model's predictions, quantifying the significance of each influencing factor on Ldam longevity and providing reference for engineering professionals. Moreover, collaborations were conducted with technical personnel from a local enterprise in Yangquan City, Shanxi Province, where on-site investigations were carried out and first-hand data of 46 sets of Ldam were obtained using drone-based 3D laser scanning technology. These data further enhance the value of this study, enrich the currently scarce database of Ldam, and further validate the generalization ability of the model and its effectiveness in future practical engineering applications. In conclusion, this study enriches the currently scarce database of Ldam and provides an effective method for accurate longevity prediction, demonstrating significant practical significance.