Development of a Hybrid SG-ARIMA-LSTM Model for Improved Meteorological Drought Forecasting Using the Standardized Precipitation Index in the KwaZulu-Natal Province, South Africa

Abstract Drought is a significant natural hazard that has devastating effects on both human life and water resources. Monitoring and predicting drought are essential for the efficient management of water resources, thereby reducing its effects. Designing a consistent drought prediction model based on the dynamic relationship between the drought index and its prior values remains difficult due to nonstationarity and nonlinearity. This study investigates the combined strengths of the Savitzky–Golay (SG) filter, autoregressive integrated moving average (ARIMA), and long short-term memory (LSTM) to test a new method of a hybrid model’s ability to accurately forecast future droughts in uMkhanyakude District, South Africa, using standardized precipitation index (SPI) as a drought assessment. The SPI was computed for 6-, 9-, and 12-month time scales using monthly rainfall data from 1980 to 2023 (528 monthly observations) for a 44-yr period. The performance of the models is evaluated using three statistical measures, namely, root-mean-square error (RMSE), directional symmetry (DS), and coefficient of determination ( R 2 ). The results reveal the SG–ARIMA–LSTM hybrid model as an efficient tool, outperforms the individual models, SG–ARIMA, SG–LSTM, and ARIMA–LSTM in forecasting across all time scales with the improved RMSE ranging between 0.2056 and 0.3011 for SPI-6, 0.1051–0.2064 for SPI-9, and 0.0525–0.0854 for SPI-12, and the R 2 ranges between 0.9392 and 0.9624 for SPI-6, 0.9724–0.9865 for SPI-9, and 0.9904–0.9969 for SPI-12. The SG–ARIMA–LSTM-based approach proposed herein could be adopted to forecast the drought with reasonable accuracy. Significance Statement This study introduces a novel hybrid modeling approach that integrates the Savitzky–Golay filter, autoregressive integrated moving average (ARIMA), and long short-term memory (LSTM) networks to address the complexities of nonstationarity and nonlinearity in meteorological drought prediction. By employing the standardized precipitation index (SPI) across various time scales, the proposed model exhibits enhanced forecasting accuracy relative to individual and pairwise models. The results highlight the potential of the Savitzky–Golay (SG)–ARIMA–LSTM model as a robust instrument for improving drought prediction in the uMkhanyakude District, South Africa, thereby contributing to more effective water resource management and the development of early warning systems in vulnerable regions.