Precipitation forecasts are of high significance for different disciplines. In this study, precipitation was forecasted using a wide range of teleconnection signals across different precipitation regimes. For this purpose, four sophisticated machine learning algorithms, i.e., the Generalized Regression Neural Network (GRNN), the Multi-Layer Perceptron (MLP), the Multi-Linear Regression (MLR), and the Least Squares Support Vector Machine (LSSVM), were applied to forecast seasonal and annual precipitation in 1- to 6-months lead times. To classify precipitation regimes, precipitation was clustered using percentiles. The indices quantifying El Niño-Southern Oscillation (ENSO) phasing showed the highest association ith autumn, spring, and annual precipitation over the studied areas. The MLP and LSSVM algorithms provided satisfactory forecasts for almost all cases. However, our results indicated that the performance of LSSVM decreased in testing data, implying the tendency of this algorithm towards overfitting. The MLP showed a more balanced performance for the training and testing sets. Consequently, MLP seems best suited to be used for forecasting precipitation in our study area. The modeling algorithms provided less reliable forecasts for the regions corresponding to the 10–40th percentiles, mostly located in hyper-arid and arid environments. This underscores the inherent difficulty of precipitation forecasting in the hyper-arid and arid areas, wherein precipitation is very erratic and sparsely distributed. Our findings illustrate that clustering precipitation regimes to consider microclimate seems vital for reliable precipitation forecasting. Moreover, he results seem useful to design preventive drought/flood risk management strategies and to improve food-water security in Iran.
