Assessment of Flood Frequency using Statistical and Hybrid Neural Network Method: Mahanadi River Basin, India

• Published in: Journal of the Geological Society of India 2021-08, Vol.97 (8), p.867-880
• Main Authors: Samantaray, Sandeep, Sahoo, Abinash, Agnihotri, Ankita
• Format: Article
• Language: English
• Subjects: Adaptive systems; Algorithms; Artificial neural networks; Discharge; Earth and Environmental Science; Earth Sciences; Environmental risk; Extreme values; Flood discharge; Flood forecasting; Flood frequency; Flood management; Flood predictions; Flooding; Floods; Frequency analysis; Fuzzy logic; Geology; Heuristic methods; Hydraulic structures; Hydrogeology; Mathematical models; Neural networks; Optimization; Peak floods; Regression analysis; River basins; River discharge; Rivers; Stations; Statistical analysis; Statistical methods
• ISSN: 0016-7622; 0974-6889
• DOI: 10.1007/s12594-021-1785-0

Flooding is the most common and widespread natural hazard affecting societies around the globe. In this context, forecasting of peak flood discharge is necessary for planning, designing and managing hydraulic structures and is crucial for decision makers to mitigate flooding risks. This study investigates potential of four most frequently used traditional statistical distribution techniques and three neural network algorithms for flood forecasting. Four statistical methods includes Generalized Extreme Value (GEV), Log Pearson-III (LP-III), Gumbel, and Normal. The methods were used for modeling annual maximum discharge at Andhiyarkore, Bamanidhi, Baronda, Kurubhatta gauge station of the river Mahanadi for a period of 60 years (1960 to 2019). In addition, a new hybrid neural network approach (ANFIS-FFA) combining the optimization model i.e. Firefly Algorithm (FFA) with data-driven model Adaptive Neuro Fuzzy Inference System (ANFIS) is adopted to predict flood discharge and compare the obtained results with conventional algorithms. Three statistical constraints MSE, RMSE, WI are employed to find the performance of proposed hybrid model. Result shows that, ANFIS-FFA gives the best values of WI as 0.9604, 0.961, 0.9598 and 0.9615 at Andhiyarkore, Bamanidhi, Baronda, Kurubhatta gauge stations respectively during testing phase. Again regression analysis is done to find the value for coefficient of determination; it gives the best value of R 2 as 95.906, 96.014, 96.113, 96.131 at Andhiyarkore, Bamanidhi, Baronda, Kurubhatta gauge stations considering ANFIS-FFA algorithm. Results from this comparative exercise suggest that hybrid ANFIS-FFA gives best performance compared to other statistical and conventional neural network approaches.