Hydrological drought forecasting using multi-scalar streamflow drought index, stochastic models and machine learning approaches, in northern Iran

Abstract

Hydrological drought is an environmental event that affects surface water resources such as surface runoff and reservoir levels and its prediction, can help the water managers to be aware of the future status of the region. This study aims to predict and evaluate hydrological drought in the southwestern margin of the Caspian Sea. For this, Streamflow Drought Index (SDI) is used as a multi-scalar hydrological drought indicator and calculated in time windows of 1, 3, 6, 9 and 12-month. In this study, time series stochastic models were used for the first time to predict SDI and compared with two black-box machine learning (ML) methods including Adaptive Neuro-Fuzzy Inference System (ANFIS) and Group Method of Data Handling (GMDH). The used data belongs to two rivers named Khalkaei and Pasikhan in Guilan province, and the period of 1986–2015 on a monthly scale. Autocorrelation and Partial Autocorrelation Functions (ACF and PACF) was used for selecting inputs among the series’ monthly time lags. The results showed that there are seasonal trends in SDI’s 1, 3, 6 and 9-month time windows; therefore, these time windows have better adaptability with the Seasonal Autoregressive Integrated Moving Average (SARIMA) model. But there is no seasonal trend in the 12-month time window (SDI12) series and the non-seasonal Autoregressive Moving Average (ARMA) model was found as the best choice for this time window. Among the models, the MLs (GMDH and ANFIS) had approximately similar prediction accuracies. Comparing the models indicate that the linear stochastic models in addition to the simplicity of use were significantly more accurate than the complex non-linear ML models. The current results suggest using the stochastic models for hydrological drought forecasting, in catchment areas.