HydroBoost Hybrid Model: A Novel SARIMAX–GBM Residual Learning Approach for Rainfall Spike Prediction

Accurate rainfall forecasting is necessary for agriculture, water management and, disaster and drought preparedness particularly in semi-arid regions like Ahmedabad, Gujarat, India. The forecasts are considered reliable not only if it just captures the overall trends and seasonal patterns but also sudden rainfall spikes and irregular variations. This study proposes a hybrid forecasting model that combines Seasonal AutoRegressive Integrated Moving Average (SARIMAX) with the rolling mean and Gradient Boosting Machine (GBM) with moving average to overcome the limitations of traditional models in learning nonlinear variations in the rainfall data. The data utilized in this research is the monthly resampled data of Ahmedabad, ranging from 1969 to 2021, obtained by the IMD. A hybrid approach is used, where the SARIMAX component models long-term trend and seasonality, and the GBM component learns from its residuals to model complex behaviour. This hybrid approach helps in improving accuracy of the forecast as well as learn the nonlinearity better, which poses as a limitation in the classical models. Along with the forecasting, a comparative analysis has also been carried out to highlight the better results obtained for the hybrid model in comparison to the other base models. The model's dual nature ensures that it learns the rainfall peaks well enough, while the classification metrices helps to know the accuracy of the model and visualizations helps support the overall forecast validity of the model. Overall, the proposed method helps provide an effective solution for forecasting rainfall in the current rapidly changing climate.