Projection of temperature and precipitation for 2020-2100 using post-processing of general circulation models output and artificial neu-ral network approach, case study: Tehran and Alborz provinces

Multi-model projections in climate studies are performed to quantify and narrow uncertainty and improve reliability in climate projections. The challenging issue is that there is no unique way to obtain performance metrics, nor is there any consensus about which method would be exactly the best method for combining models. The goal of this study was to investigate whether combining climate model projections using an artificial neural network approach could improve climate projections and therefore reduce the range of uncertainty. The equally-weighted model averaging (the mean model) and single climate model projections (the best model) were also considered as a reference of comparison for our artificial neural network combination approach. Simulations of historical climate and future projections from 15 General Circulation Models for temperature and precipitation were employed.Our results indicate that based on calculated performance indices combining General Circulation Models projections by using the artificial neural network approach significantly improves the simulations of temperature and precipitation for the historical period compared to the best model approach and the mean model approach. Our results also indicate that based on the calculated performance indices for the three approaches, projections based on single model simulation might not yield reliable results because the best model changed between temperature and precipitation, and also among stations that were studied. Therefore, there was no a unique model which could represent the best model for all climate variables and/or stations in the study region. The mean model was also not skillful enough in giving an accurate projection of historical climate compared to the other two approaches. Therefore, the ANN approach was used to estimate projections of future temperature and precipitation for the study region based on three different emission scenarios.Simulation of temperature indicated that the artificial neural network approach had the best skill at simulating monthly means of the historical period compared to other approaches in all stations. Simulation of precipitation in the historical period, however, indicated that the artificial neural network approach was not the best approach in all stations, although this modeling approach performed better than the mean model approach. Multi-model projections of future climate variables for this study region performed by the artificial neural network approach projected an increase in temperature and reduction in precipitation in all stations and for all scenarios.The artificial neural network approach can benefit projections of the climate variables and has the potential to reduce the uncertainty aspects in constructing and combining metrics used for weighting the models. However, this approach is subject to some limitations which exist in similar skill-based performance studies of models and should be considered in future similar studies.