Performance Analysis of NAR Model for Short and Long Term Drought Forecasting in Arid Regions

Drought forecasting is of particular importance in water resources management. Drought forecasting allows planners to schedule for reducing the negative impacts of drought as well as to adapt to it. Drought prediction is more important in arid regions. Because these areas are inherently water scares and the consequences of drought in these areas are more severe. Due to the high variabilities of the temporal and spatial distribution of precipitation in these areas, the frequency of drought is higher and results in more difficulty to model and predict drought. In this study, since drought time series is nonlinear and cyclic, nonlinear autoregressive neural networks (NARs) were used to predict short-term and long-term drought in Yazd synoptic station from 2006 to 2018. Reconnaissance Drought Index (RDI) which in addition to precipitation, considers potential evapotranspiration to monitor droughts, for one, three, and six months timescales was calculated. Potential evapotranspiration was calculated using the FAO-Penman-Monteith method. The results of short-term (one month) drought prediction presented that the model provides high performance in predicting three and six months RDI values. The results of long-term (13-years) drought forecasting (without access to real drought data from 2006 to 2018) indicated that RDI values in dry months show best fit to real values ​​ in ​​ three months’ timescale. To improve the efficiency of the model in the long-term drought forecasting, long-term precipitation and potential evapotranspiration (without model access to real data from 2006 to 2018) were predicted. RDI values ​​were then calculated based on the predicted precipitation and potential evapotranspiration data. The results showed that the prediction accuracy increased in one and three months scales. Also, on six months’ timescale, RDI data were more accurately predicted in dry months.