Modeling Precipitation: A Statistical and Machine Learning Approach

Abstract

Precipitation modeling is indispensable for understanding hydrological processes, optimizing water resource management, and mitigating precipitation-induced geophysical mass flows such as debris and mud flow risks. This study employed both statistical and machine learning techniques to model precipitation patterns in Kathmandu, Nepal. Specifically, it compares the performance of the autoregressive integrated moving average (ARIMA) model and the Support Vector Regressor (SVR), using five years of historical meteorological data from the Department of Hydrology and Meteorology, Babarmahal. Stationarity tests confirmed the dataset's stationarity, allowing the use of ARIMA without differencing. Initial ARIMA parameters were selected using ACF and PACF plots and refined through grid search with AIC scores. Performance evaluation demonstrated that SVR outperformed ARIMA in both in-sample and out-of-sample predictions, as indicated by lower RMSE and MAE and higher Nash-Sutcliffe Efficiency (NSE) values, suggesting a better data fit and generalization. Furthermore, the Durbin-Watson statistic confirmed SVR's superior handling of autocorrelation, reinforcing its effectiveness in precipitation forecasting. However, both models faced challenges in accurately predicting extreme precipitation events. Future research should explore a hybrid approach integrating statistical models, machine learning, and deep learning to enhance predictive performance. Additionally, further consideration of meteorological and geographical parameters may improve model accuracy. These findings can support meteorologists, urban planners, disaster management authorities, and policymakers in strengthening early warning systems, improving water resource management, and enhancing resilience against climate-related challenges.