Adaptive Modeling for Rainfall Prediction Using Ensemble Machine Learning and Bayesian Optimization

Heavy rainfall in Indonesia will usually be followed by news of a flood disaster. Flood is a natural phenomenon that is detrimental to social life. There are various approaches to studying rainfall to forecasting floods disaster, one of which is ensemble machine learning. Ensemble machine learning produces better performance models than single machine learning when analyzing the occurrence of floods through rainfall. However, one model does not seem sufficient to analyze the overall occurrence of rainfall in different areas. This happens because the datasets in each region will have different characteristics, so the models generated by machine learning need to be distinct from one another. Furthermore, manual tuning in ensemble machine learning is sometimes difficult due to a large number of hyperparameters, the complexity of a given problem domain, and ineffective model evaluation. This paper proposes an adaptive modeling approach for rainfall prediction based on these problems using stacking ensemble machine learning Bayesian Optimization. The mean square error (MSE) result of Surabaya, Bandung, Jakarta, Semarang, and Banten are 14.755, 0.249, 3.310, 0.243, and 0.917 respectively. The results of the study indicate that this research provides a more dynamic and reliable approach to determining the best tuning model to predict rainfall. We provide some models in one time running with the tunning configuration in each dataset dynamically.