Evaluation of stochastic and artificial intelligence models in modeling and predicting of river daily flow time series

Modeling and forecasting the flow of rivers, especially in flood-prone areas using warning systems, enables officials to take the required measures for cutting the damage. On the other hand, they can adopt specific measures flood control and prevention. In the present study, two stochastic and three artificial intelligence (AI) models were compared, in modeling and predicting the daily flow of the Zilakirud river in northern Iran. The daily data belongs to the period of 2001–2015 (14 hydrological years from 23/Sep/2001 to 22/Sep/2015). First, the data was reviewed in terms of hydrological drought at the annual scale, using Streamflow Drought Index (SDI). The inputs for the models included the time lags of river daily flow. After choosing the input scenario, two approaches were tested for choosing the percentage of calibration and validation: (1) The last single year for validation; (2) The last 4 years for validation (about 30% of the data, which is a common method). A comparison between the models showed that the accuracy of AI models was higher than stochastic ones. Among the AI models, Group Method of Data Handling (GMDH) and Multilayer Perceptron (MLP) showed the best validation performance in both approaches. The findings showed that among the two approaches, approach (1) can show a better predicting accuracy with RMSE of 1.50 and 1.40 CMS for GMDH and MLP, respectively while in the second approach, the RMSE was 5.15 and 5.29 CMS for GMDH and MLP, respectively. Also, from the perspective of drought classes, the weakest result belonged to the moderately wet hydrological year (the hydrological year of 2011–2012) and the best performances was observed in the mild drought hydrological year (the hydrological year of 2014–2015).