Predicting rainfall amount is essential in water resources planning and for managing structures, especially those against floods and long-term drought establishment. Machine learning techniques can produce good results using a minimum dataset requirement, making it a leader among the prediction algorithms. This work develops a hybrid learning model for monthly rainfall prediction at four geographical locations representing Mediterranean basins in Northern Algeria and desert areas in Egypt. The study proposes an adaptive dynamic-based hyperparameter optimization algorithm to improve the accuracy of hybrid deep learning models. The proposed model provided a good fit, based on the obtained Nash-Sutcliffe efficiency index (NSE ≈ 0.90) with a high correlation coefficient of R ≈ 0.96, providing improvements of up to 62% in the RMSE. The proposed method proved to be an encouraging and promising tool to simulate water cycle components for better water resources management and protection.