A Comparative Analysis of Rainfall Forecasting in Cox’s Bazar using SARIMA, LSTM and GRU

Evaluating climate change, particularly rainfall fluctuations, is critical for regions prone to extreme weather, like Cox's Bazar, Bangladesh. This study employs both traditional time-series models and deep learning techniques to forecast rainfall patterns. An Exploratory Data Analysis (EDA) was performed on the monthly average rainfall data (1981-2022) from the Bangladesh Meteorological Department, revealing significant trends and seasonal patterns that guided model selection. ARIMA (0,0,1)(2,0,1)12, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) models were developed to predict future rainfall. Among them, the GRU model achieved the lowest training errors (RMSE: 5.61, MAE: 3.48), while LSTM excelled in test data performance, producing lower errors (RMSE: 7.79, MAE: 4.68). Additionally, LSTM demonstrated the smallest percentage difference between training and testing errors, highlighting its superior generalization capabilities compared to SARIMA and GRU. These findings underscore the effectiveness of LSTM in capturing complex rainfall patterns, contributing to improved flood risk assessment and water resource management strategies.