Assessment of machine learning model performance for seasonal precipitation simulation based on teleconnection indices in Iran

Precipitation is one of the most important factors affecting the climate, hydrological processes, and living environment. Hence, the precipitation forecast is significant for water resource exploitation and preparation for extreme climatic events such as drought and floods. In this context, teleconnection indices are commonly used as predictors across the globe. However, most studies have focused on investigating the correlation between seasonal precipitation and teleconnection indices from the meteorological stations by using some limited models to simulate the precipitation. This study evaluated the use of 40 teleconnection indices by exploiting 4 machine learning models (ML), namely generalized regression neural network (GRNN), multi-layer perceptron (MLP), least squares support vector machine (LSSVM), and multilinear regression (MLR) to forecast and model seasonal precipitation in a larger scale than meteorological stations, specifically main basins, and sub-basins of Iran. For that purpose, the seasonal precipitations in 6 main basins, including 30 sub-basins, were selected based on 717 stations for the period 1987–2015. First, the correlations between 40 teleconnection indices and the seasonal precipitation of sub-basins were measured by Pearson correlation to determine their significance using a correlation matrix. Then, the most significant (predictor) variables with time lags of 1 to 6 months (for each season) were extracted by a stepwise procedure per sub-basin and considered as input for the four ML models. Finally, the performances of the models were assessed based on coefficient of determination ( R 2 ), root mean square error (RMSE), mean bias error (MBE), and scatter index (SI) statistic tests. According to the results, the most significant correlation between teleconnection indices and autumn, winter, and spring precipitations occurred with time lags of 1–4, 4–6, and 1–4 months, respectively, in some teleconnection indices. Evaluation of the simulation, the LSSVM model, performed excellently followed by MLP for most of the sub-basins, while GRNN and MLR models showed a poor simulation performance between 1987 and 2015. The results showed that the LSSVM had more accuracy and less RMSE in the training period than other models, while the MLP and GRNN models’ RMSE were less than the MLR and LSSVM. Therefore, the MLP and GRNN are recommended in modeling and forecasting seasonal precipitation in Iranian sub-basins. The overall results confirmed t