Evaluating Seasonal Variability in Rainfall Forecast Accuracy Using Comparative Regression Modelling Approaches

This study compares the seasonal variation in the accuracy of meteorological forecasting of monthly and annual rainfall (mm) using ten conventional regression and machine learning models. The January to December meteorological data provided monthly on the Andaman Islands were reviewed on rainfall, temperature, humidity, and wind speed between 1950 and 2022. The performance of multiple predictive models, such as Ridge, Lasso, Elastic Net, Random Forest, Gradient Boosting Machine (GBM), LSTM, CNN, SVR and XGBoost, were implemented and evaluated based on RMSE, MAE, MAPE and R² metrics. Results show that there are significant seasonal variations in model performance. The comparative analysis reveals significant seasonal fluctuations in model precision. Specifically, SVR achieved the highest accuracy in winter (DJF) with an RMSE of 72.31, while XGBoost performed optimally in spring (MAM) with an RMSE of 93.2 and an R² of 0.76). The traditional regression models had moderate performances but were weaker in seasons of high variability. These findings demonstrate that it is of utmost importance to use season-related models and clearly specify the target variables to maximize the meteorological forecasting and climatic-based planning of water resources.