An Adaptive Neuro‐Fuzzy Inference System ( <scp>ANFIS</scp> ) and a Wavelet‐Integrated <scp>ANFIS</scp> ( <scp>WANFIS</scp> ) for Univariate Bias‐Correction of <scp>GCM</scp> ‐Simulated Temperature

ABSTRACT Correcting systematic errors or biases of raw outputs of a global climate model (GCM) simulating a climatological variable is an important requirement for the reliable use of these outputs in climate change impact assessments. In this study, a machine learning algorithm, namely Adaptive Neuro‐Fuzzy Inference System (ANFIS), was first used for univariate bias correction of GCM‐simulated outputs of daily maximum, mean, and minimum temperatures by considering one of these variables at a time. The ANFIS was then integrated with a discrete wavelet transform (DWT) in devising a novel bias‐correction technique, named WANFIS, wherein (i) high‐ and low‐frequency components of a time series of a temperature variable were first produced by DWT of that time series up to a pre‐determined level of resolution, (ii) subsequently, an ANFIS was separately applied to each of the high‐ and low‐frequency components of raw GCM‐simulated data for correcting bias in relation to the corresponding components of the concurrent reference data and (iii) the bias‐corrected components were finally aggregated by using DWT again to reconstruct the bias‐corrected time series of the selected temperature variable. The performances of ANFIS and WANFIS were compared with that of a traditional univariate technique, quantile delta mapping (QDM), for correcting bias. The techniques were applied to gridded outputs of GCM‐simulated temperature variables over diverse physiographic and climatic regions across mainland India. The ERA5 reanalysis data sets produced by the European Centre for Medium‐Range Weather Forecasting were used as reference data for investigating relative performances. The WANFIS emerged as being an efficient bias‐correction technique that consistently outperformed the ANFIS and QDM techniques in simulating spatiotemporally averaged as well as spatially distributed temperature variables, whereas some inconsistencies in performance were noted in the case of ANFIS and QDM. The discernible spatial patterns in the variation of performance measures under distinct physiographic and climatic conditions were also relatively uniformised towards higher levels of performance on application of WANFIS.