An Explainable Ensemble Machine Learning Model for Short-Term Flood Occurrence Prediction Using Hydro-Climatic Time-Series Data

Floods are really harmful surprise elements of nature, and as such, making short-term flood prediction very accurate is a major requirement of early-warning systems done in such a way as to prevent human and property losses, economic disruption, etc. Still, they are hard to guess due to the very intricate combination of hydrological, climatic, and other environmental factors. The present paper offers an interpretable ensemble machine learning framework for predicting the times of flood events around 1-3 days ahead via the use of hydrology, climate, and environmental indicators together. The method offers great help through data preprocessing, feature normalisation, and the application of various regression models to cost-continuous flood probabilities estimation. Random Forest and Gradient Boosting algorithms are used to find and improve prediction accuracy through capturing non-linear relationships, while a hybrid ensemble method combines the advantages of individual models. Decision-making is made simplistic by the conversion of probabilistic outputs into binary flood alerts at a predetermined fixed threshold. The framework is executed and tested in a MATLAB-Simulink setting, and the analysis confirms its readiness for real-time operations. The results from the experiments indicate that the learning through the ensemble approach has significantly improved the prediction reliability and interpretability as compared to single model techniques.