Comparison of Artificial Neural Network, Decision Tree and Bayesian Network Models in Regional Flood Frequency Analysis using L-moments and Maximum Likelihood Methods in Karkheh and Karun Watersheds

Abstract

Proper flood discharge forecasting is significant for the design of hydraulic structures, reducing the risk of failure, and minimizing downstream environmental damage. The objective of this study was to investigate the application of machine learning methods in Regional Flood Frequency Analysis (RFFA). To achieve this goal, 18 physiographic, climatic, lithological, and land use parameters were considered for the upstream basins of the hydrometric stations located in Karkheh and Karun watersheds (46 stations with a statistical length of 21 years). The best Probability Distribution Function (pdf) was then determined using the Kolmogorov-Smirnov test at each station to estimate the flood discharge with a return period of 50-year using maximum likelihood methods and L-moments. Finally, RFFA was performed using a decision tree, Bayesian network, and artificial neural network. The results showed that the log Pearson type 3 distribution in the maximum likelihood method and the generalized normal distribution in the L moment method are the best possible regional pdfs. Based on the gamma test, the parameters of the perimeter, basin length, shape factor, and mainstream length were selected as the best input structure. The results of regional flood frequency analysis showed that the Bayesian model with the L moment method (R2 = 0.7) has the best estimate compared to other methods. Decision tree and artificial neural network were in the following ranks.

Keywords

• modeling
• instantaneous maximum discharge
• machine learning models
• l-moments
• bayesian network