Search

Changes in patterns of meteorological parameters, like precipitations, temperature, wind, etc., are causing significant increases in various extreme events. And these extreme events, i.e., floods, heatwaves, hurricanes, droughts, etc., lead to a shortage of water resources, crop failures, wildfires, and economic losses. However, Global Circulation Models (GCMs) are considered the most important tools for quantifying climate change. Therefore, we selected 20 different GCMs of precipitation in our research, as the frequency of extreme events, like drought and flood, is highly related to changes in precipitation patterns. However, this research introduced a new weighting scheme — MCFSAWS-Ensemble: Monte Carlo Feature Selection Adaptive Weighting Scheme to Ensemble multiple GCMs, whereas, Monte Carlo Feature Selection (MCFS) is one of the most popular algorithms for discovering important variables. However, the proposed weighting scheme (MCFSAWS-Ensemble) is mainly based on two sources. Initially, it evaluates the prior performance of each GCM model to define their relative importance using MCFS. Then, it computes value by value difference between the observed and simulated model. In addition, the application of this paper is based on the monthly time series data of precipitation in the Tibet Plateau region of China. In addition, we used twenty GCMs from the Coupled Model Intercomparison Project Phase 6 (CMIP6) to analyze the implications of the MCFSAWS-Ensemble. Further, we compared the performance of the MCFSAWS-Ensemble scheme with Simple Model Averaging (SMA) through Mean Average Error (MAE) and correlation statistics. The results of this research indicate that the proposed weighting scheme (MCFSAWS-Ensemble) is more accurate than the SMA approach. Consequently, we recommend the use of advanced machine learning algorithms such as MCFS for making accurate multi-model ensembles. [ABSTRACT FROM AUTHOR]

Artificial neural networks are widely applied for different forecasting applications including rainfall forecasting. The climate input features and parameters for neural networks highly affect the overall performance of the prediction model. Therefore, an appropriate approach for the selection of features and parameters is needed. In this paper, a novel approach is proposed to select the input features and neural network parameters. A hybrid genetic algorithm that combines natural reproduction and particle swarm optimization characteristics was developed to select the best input features and network parameters for each month. The developed model was compared against alternative climate and network parameters feature selection model, climate feature selection model and climatology where a better accuracy was recorded with the proposed model. The skill score against the three alternative climate models was 17.41, 21.68 and 32.12%, respectively. The aggregated time series of the proposed model showed a root-mean-square error of 141.67 mm for a location with 3553.00 mm annual average. [ABSTRACT FROM AUTHOR]