Abstract

Floods are unpredictable natural disasters, often triggered by extreme weather conditions, particularly in countries like India. This study utilises machine learning to predict runoff, with a focus on attention-based models, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures. We compare Attentional LSTM, PCA-based LSTM, Self-Attention LSTM, and Self-Attentional GRU models across various window lengths (1, 7, 14, and 30 days). The Attentional LSTM model consistently outperforms others, achieving the lowest Mean Squared Error (MSE=0.000476) and Mean Absolute Error (MAE=0.015475) while attaining the highest Nash-Sutcliffe Efficiency (NSE=0.978). A 14-day window size provides the best balance between predictive accuracy and computational efficiency. While other attention-based models, such as the Self-Attentional GRU (MSE=0.000654, MAE=0.018878, NSE=0.970), perform well, simpler models without attention mechanisms show significantly lower accuracy. Additionally, incorporating Principal Component Analysis (PCA) does not enhance performance, as the PCA-Based Attention LSTM (MSE=0.000855, MAE=0.022152, NSE=0.961). The model underperforms compared to the Attentional LSTM. Explainable AI (XAI) techniques such as LIME are used to analyse model behaviour, highlighting the importance of attention mechanisms in improving time-series predictions for hydrological applications.