Flood is one of the most important and destructive natural hazards that endangers the life, economic, social, environmental and psychological security of the earth's inhabitants every year. The pattern of this phenomenon has changed in recent years due to many factors such as climate change. Therefore, flood forecasting and risk assessment of this phenomenon under the factor of nonstationry such as climate change can play a vital role. For this purpose, in this study, the eva‎luation of two important goals, including flood forecasting and flood risk assessment in the watershed of the Ken River, located in the north of Tehran province, has been done as a case study. Due to its location and special conditions, this basin is considered one of the important and critical basins for the study of flood behavior. In the first phase of this study, a new fusion framework for flood forecasting based on machine learning, statistical and geostatistical models has been introduced. For this purpose, at first, a machine learning model was used to fill the gap of ground observation data of the historical period based on remote sensing precipitation data including ERA5, CHIRPS and PERSIANN-CDR. Then, four individual machine learning models, including Random forest, Multi-layer perceptron, Support vector machine and Extreme learning machine, are developed to simulate the daily flow on a long-term scale. In the next step, three hybrid models, including Random forest, Bayesian model averaging, and Bayesian maximum entropy, have been used to combine the outputs of the individual machine learning models of the previous step and to improve the results, especially in the prediction of further steps. The proposed framework has also been implemented for daily variables downscaling of three climate general circulation models (GCMs) under two emission scenarios. The results of flood forecasting step showed that the individual models demonstrate weak performance, especially in predicting the daily flow of long time steps, so it is necessary to use a fusion technique to improve the results. Also, in this step, the random forest model has shown high efficiency in the fusion step compared to other fusion models. On the other hand, this technique has also shown an effective performance in the biased correction of daily precipitation data of GCMs. In the second step of the research, the flood risk of the study area has been eva‎luated under different climate change models and scenarios based on Copula's approach. As we know, flood is a multidimensional event and requires the analysis of many factors. Most of the research often uses a bivariate framework relying on historical data, while the bivariate flood approach can lead to a decrease in estimation accuracy, especially as climate change is expected to affect flood frequency analysis and flood system design in the future. Therefore, the second step of this study is to eva‎luate the predicted changes in three important characteristics of floods, including duration, flood volume and flood peak, and the joint return period of floods using copula functions based on eight general circulation models. In this phase of the study, analysis has been done based on two emission scenarios including SSP2-4.5 and SSP5-8.5 for three 31-year periods including the far future (2070-2100), the mid-term future (2070-2040), and the historical period (1982-2012). In this step, the outputs of precipitation and minimum and maximum temperature of biased corrected climate change models have been used as predictors of the machine learning model for daily flow simulation. Then, a trivariate framework based on copula functions has eva‎luated the trivariate flood frequency. These analyzes have been presented using hierarchical Archimedean copula in three structures including symmetric, asymmetric homogeneous and asymmetric heterogeneous copula, and finally their performance in estimating flood frequencies has been eva‎luated.

حميدرضا صفوي , محمد رضا نيكو

محمدرضا نجفي

محمدحسين گل محمدي , رامتين معيني , ناصر طالب بيدختي