Use of stepwise m5 model tree to forecast the P24max based on teleconnection indices

Abstract

In this study, the linear and non-linear multivariate relationships between 25 teleconnection indices (tele-indices) as independent variables and annual P24max as the dependent variable were analyzed using multivariate linear regression (MLR) and decision tree regression models (M5), in selected synoptic weather stations of Iran over a statistical period of 30 years (1992–2021). No strong and statistically significant correlation between each tele-index and P24max was observed. Therefore, it is not appropriate to attribute climate changes in the region to a single factor such as El Niño, but rather consider the combined influence of multiple factors. The M5 model demonstrated higher performance, indicating a non-linear relationship between tele-indices and P24max. The stepwise execution of the M5 model tree showed that the algorithm follows a greedy approach, and it is not necessary to use all variables to predict P24max. The normalized root mean square error (NRMSE) of P24max estimation was found to be 15%, 13%, 15%, 8%, 20%, 14%, and 12% with the coefficients of determination of 0.78, 0.79, 0.72, 0.85, 0.81, 0.82, and 0.84 in Hashemabad-Gorgan, Rasht, Kermanshah, Ahvaz, Bandar Abbas, Isfahan, and Birjand, respectively. Finally, it is possible to forecast P24max using tele-indices measured in the previous year.