SMRF: a new class-based probabilistic approach for season-ahead monthly rainfall forecasting

Abstract

Monthly rainfall forecasting is an important task in hydrology. Because of the stochastic nature of rainfall events, probabilistic analysis is considered an appropriate approach for rainfall forecasting. This article introduces a new probabilistic hybrid model, called SMRF, for season-ahead monthly rainfall forecasts. The SMRF model is based on a combination of classification and probabilistic kernel function and conducted by a four-step algorithm. It benefits from probabilistic kernel functions calculated for each of humid classes of seasonal rainfall in each of months of the season. This model is favoured for monthly rainfall forecasting due to its ability to simultaneously estimate rainfall for all months within a season, without relying on the rainfall patterns of preceding months. The new model was demonstrated using different periods of rainfall data from two different climate divisions, Karkheh (1982–2014) and Kardeh (1993–2020) basins, in southwest and northeast of Iran. The efficiency of the proposed model was verified by comparing two probabilistic artificial intelligence models, namely Generalized Regression Neural Network (GRNN) and K-Nearest Neighbour model (KNN), based on the K-fold cross-validation method. The results showed that the SMRF was superior to the GRNN and KNN in both study areas, while in most cases, the accuracy of the SMRF was higher than those of the benchmarks, particularly for heavy rainfall forecasting. Considering the seasonal average of Nash–Sutcliffe efficiency criteria, the SMRF showed up to 170% (4%) and 100% (80%) more accurate forecasts than the GRNN and KNN in Karkheh (Kardeh) basin, respectively.