Data-driven ensemble model for probabilistic prediction of debris-flow volume using Bayesian model averaging: Fid-Bau Portal

Data-driven ensemble model for probabilistic prediction of debris-flow volume using Bayesian model averaging

Tian, Mi / Fan, Hao / Xiong, Zimin / Li, Lihua

Abstract Accurate and reliable predictions of the debris-flow volume are the necessary prerequisite for potential hazard delineation and risk assessment of debris flows. Various theoretical, empirical, and machine learning methods have been proposed by researchers to estimate the debris-flow volume. However, current methods generally provide point-value deterministic predictions and have limitation in assessing the predictive uncertainties associated with the observation data, model parameters, and structures. This paper proposed a data-driven ensemble model to probabilistically forecast the debris-flow volume using multiple deterministic machine learning methods and Bayesian model averaging (BMA). The rainfall-induced debris flows in Taiwan were selected as an illustrative example to evaluate the feasibility of the proposed approach. Firstly, the debris-flow datasets are preprocessed by the principal component analysis (PCA) to select input variables. Then, four data-driven models are applied to provide deterministic estimates for ensemble forecasts. Finally, BMA incorporates the deterministic predictions of multiple data-driven models to generate probabilistic forecasts. The performances of individual data-driven models and BMA ensemble forecast are evaluated and compared. Results show that the proposed BMA ensemble model performs better than the single models for predicting the debris-flow volume in terms of the effectiveness and robustness. Ensemble models with good performance can combine the strengths of different models to improve the prediction accuracy. Weighting only good members may not achieve the best performance for both calibration and validation periods. The performance of different combinations of data-driven models is closely related to the observation data and the prediction accuracy of each model.