Abstract
Extreme rainfall event forecasting is important as these events are responsible for causing floods, landslides, and other hazards with substantial socio-economic consequences. The intricate nature of rainfall occurrences makes it more difficult to forecast accurately, especially when it comes to extreme rainfall. This study investigates the application of the Quantile Perturbation Method (QPM) along with the Long Short-Term Memory (LSTM) networks to forecast extreme rainfall anomalies. This methodology utilizes the strength of QPM to decipher oscillations in time series of extreme rainfall to identify anomalies, which are then forecasted using LSTM. The model was developed for the Ajay River basin as a case study based on historical rainfall data from 1901–2022. To determine the best model, several experiments with various configurations were conducted. Performance metrics such as Nash-Sutcliffe efficiency (NSE), correlation coefficient (R), and root mean square error (RMSE) were utilized for model evaluations. The QPM-LSTM model was compared against other combined machine learning models, including Artificial Neural Networks (ANN) and Support Vector Regression (SVR). The investigation demonstrated satisfactory predictive performance by the QPM-LSTM model, achieving NSE, R, and RMSE values of 0.87, 0.93, and 7.26, respectively. Compared to the other evaluated models, these results highlighted the potential of the QPM-LSTM model as a valuable tool for forecasting extreme rainfall anomalies, offering significant benefits for water resource management and other sectors vulnerable to extreme rainfall events.