The application of Bayesian model averaging based on artificial intelligent models in estimating multiphase shock flood waves

Abstract

The multiphase shock wave phenomenon is significantly affected by accumulated upstream sediment deposition and downstream hydraulic conditions. There is a lack of studies evaluating the efficacy of intelligent models in representing multiphase debris flooding over initially dry- or wet-bed tail-waters, or over downstream semi-circular obstacles. To address this, we propose a novel methodology based on Bayesian Model Averaging (BMA), which combines predictions of three individual intelligent models [i.e., “Multi-layer Perceptron” (MLP), “Generalized Regression Neural Network”, and “Support Vector Regression”]. The models were developed through experimental study whereupon high-quality sediment depths and water levels data (n = 9000) were collected from 18 shock wave scenarios with various initial conditions in channel up- and down-stream. Experimental data and related original videos are created accessible in an online repository may be used in other researches. Each model’s results were in close concord with the experimental data; RMRE and RMSE values were in the range of 1.54–5.99 mm and 1.21–40.49 mm, respectively (0.5–2% and 0.4–13.5%) with the MLP model marginally outperforming the other intelligent models. Based on statistical error indices, the BMA model had the best performance (up to 40% better) in estimating most data classes, and was more efficient than the best intelligent model signifying that the proposed methodology is explicit, straightforward, and promising for real-world applications.