CUCKOO SEARCH OPTIMIZATION NEURAL NETWORK MODELS FOR FORECASTING LONG-TERM PRECIPITATION

It is more crucial than ever to make quantitative prediction patterns of precipitation due to climate change and global warming concerns. The foundation for many climate change simulations is global circulation models (GCMs). However, to create finer models for regional use, researchers have been employing various downscaling strategies due to their coarse resolution. Technological developments in metaheuristic algorithms have introduced a different method for downscaling. This paper presents the application of a novel optimization algorithm, Cuckoo Search Optimization (CSO), to train feedforward neural networks to forecast long-term precipitation using three climate models, namely HadCM3, ECHAM5, and HadGEM3‐RA. The selected study area is Kuching City, Sarawak, Malaysia, and the models' performance was assessed using historical precipitation data validation through the square root of the correlation of determination (r), mean absolute error (MAE), root mean square error (RMSE), and Nash and Sutcliffe coefficient (E). With a setup of 20 nests (n), an initial alien egg-finding rate (Pa) of 0.6, 100 hidden neurons (HN), 1000 iterations (IN), and a learning rate (LR) of 1, the results demonstrated that the Cuckoo Search Optimization Neural Network (CSONN) is capable of forecasting precipitation with confidence levels of 95%~99% for r and 85%~94% for E, alongside lower RMSE and MAE. Future precipitation forecasts revealed that the city would experience an increase in mean monthly precipitation of 2%~26% in the 2030s, 0%~34% in the 2050s, and 4%~43% in the 2080s during wet seasons, relative to the 1970s. The findings also showed that mean monthly precipitation would decrease during dry seasons, ranging from 1%~4% in the 2030s, 1%~2% in the 2050s, and 3%~4% in the 2080s, compared to the 1970s.