Assessing the Performance of a Machine Learning-Based Hybrid Model in Downscaling Precipitation Data

Abstract

AbstractIn this paper, the performance of Statistical Downscaling (SD) for average monthly precipitation time series for Mahabad located in the north western part of Iran was quantified. The structure of the proposed approach is composed of two parts: input selection and average monthly precipitation simulation. Data from the Intercomparison Project Phase 5 (CMIP5) was chosen, and the input selection based on a Machine Learning-based (ML) method, i.e. the decision tree model (M5), was applied to select the most important predictors amongst many potential large-scale climate variables of the CMIP5 model. Next, the Random Forest (RF) model was trained, using the observed precipitation time series using the most important predictor variables, to downscale the precipitation time series of the CMIP5 model for the historical period (1992–2005) over the study area. Using the hybrid M5-RF model led to a reliable performance in the prediction of precipitation time series as the resulting Correlation Coefficient (R) and Root Mean Square Error (RMSE) values were 0.78 and 0.1 for calibration and 0.75 and 0.13 for validation, respectively.KeywordsStatistical downscalingMachine learningM5 modelRandom forestIran