Efficient Rainfall Forecasting Using Sequential Momentum-Based Artificial Neural Networks for Urban Flood Management

Abstract

Urban floods cause severe disruption in densely populated cities and damage essential public infrastructure. Rapid urbanization and weak drainage systems further increase the risk during heavy rainfall. Accurate short-term rainfall forecasting is required to support early warning and reduce losses. This research focuses on improving rainfall prediction accuracy using an Artificial Neural Network trained with a new optimizer named Sequential Momentum Gradient Descent. The study addresses the drawback of traditional gradient methods that use fixed momentum in training. A fixed momentum reduces adaptability and limits the model’s ability to learn changing weather trends. The proposed Sequential Momentum Gradient Descent adjusts momentum dynamically based on the current error rate. This adaptive process allows faster convergence and better stability for time-dependent rainfall data. The optimizer enhances learning efficiency by increasing momentum when the model progresses and reducing it during unstable learning stages. The system uses nine major weather features that represent regional variations in rainfall intensity. Experiments are carried out using weather datasets from Aminjikarai, Velachery, Mudichur, and Mahabalipuram regions in Chennai. The proposed model achieves higher prediction accuracy between 94.7 and 95.6 percent on all datasets. Training time is reduced by more than half compared to other baseline optimizers. The results show that the adaptive momentum control improves rainfall forecasting reliability under diverse climatic patterns. This approach contributes to practical urban flood management systems that depend on precise and fast rainfall prediction.