A hybrid prophet -LSTM model for drought forescasting in Machakos County.

Abstract

In Kenya, drought remains a challenge in its arid and semi-arid lands (ASALs) where Machakos County located in the Eastern region, it experiences such events every three to four years. Crop yields, livestock losses, and water shortages are highly negatively affected by these recurrent episodes. Forecasting models, including ARIMA, SARIMA, Numerical Weather Prediction (NWP), ARIMA, and SARIMA are the conventional models used to forecast drought in the region and they highly struggle to capture the complexity of drought patterns resulting to errors of more than 30% due to uneven seasonality and nonlinear climate dynamics. To overcome these limitations, this study created a hybrid forecasting model that combines strengths of both Prophet and Long Short Term Memory (LSTM) models. Daily and Monthly climate data from 2014 to 2024 with variables; humidity, precipitation, temperature and wind speed was used for model training and validation. In this framework, prophet was applied to extract long-term trends and seasonal effects, while LSTM was employed to capture nonlinear residuals variations, improving overall prediction accuracy. After analyzing the findings of the study using drought indices, it was discovered that the LSTM model with SPI-6, produced the most accurate findings with MAE = 0.3847, RMSE = 0.5410, R2 = 0.7833), which means that it explained 78.33% of the variability in drought. The Prophet-LSTM model produced great results, with MAE = 0.4788, RMSE = 0.5765, and R2 = 0.7539. It performed better than baseline Prophet model and kept the forecast stable even in the presence of data gaps. Eight drought events between 2014 and 2024 were discovered with an average duration of 2.5 months and 13-months of drought from October 2022 to October 2023. In general, the hybrid model offers a more reliable tool for forecasting drought in Machakos County and shows potential for application in other ASAL regions. The study introduces a tailored Prophet–LSTM framework for semi-arid climates, showing that seasonal decomposition combined with nonlinear learning reduces forecast errors and improves precision over traditional models. Future improvements may include integrating additional climate variables, expanding datasets, and enabling real-time model updates to strengthen drought preparedness and resource planning.