Flood forecasting using a hybrid extreme learning machine-particle swarm optimization algorithm (ELM-PSO) model

Abstract

Flood forecasting in India is carried out by the determination of the water level at flood-forecasting stations. The level forecasts are issued once water levels in a station reach a predefined warning level, which helps local authorities to determine response measures to the floods. A new approach has been explored in this paper, which involves using the mean daily gauge heights, mean daily rainfall, and the mean daily river discharge values of prior days to forecast the mean gauge heights up to 4 days in advance. These features were used as input for an extreme learning machine (ELM) regression model. The number of units in the ELM was optimized to obtain the maximum coefficient of determination using the particle swarm optimization algorithm (PSO) to create a hybrid ELM-PSO model. Gauge, rainfall, and discharge data of 4 decades from the Jenapur flood-forecasting station (Brahmani river, Odisha) and the Anandpur station (Baitarani river, Odisha) were used to create models for mean gauge height prediction. These models were then cross-validated using tenfold cross-validation, with mean-squared error (MSE) and the coefficient of determination (R-squared) as parameters for evaluation of the models. The models show promising results, with the 1-day-in-advance model having MSE 0.14 and R-squared 0.85 for Jenapur and MSE 0.23 and R-squared 0.75 for Anandpur.