<![CDATA[Assessing drought risk in equatorial urban hydrosystems presents a critical challenge for water resource management and climate adaptation. This study applies Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) deep learning models to forecast meteorological drought in Palembang, which is a low-lying, tidally-influenced metropolitan area in Indonesia, using a 30-year monthly rainfall dataset (1993–2024) from three monitoring stations. Models were developed to predict the Standardized Precipitation Index (SPI) across short-term (SPI-1), medium-term (SPI-3), and longer-term (SPI-6) timescales. While LSTM showed a marginal overall performance advantage over RNN, the study's principal finding is the significant spatial heterogeneity in drought risk across the three stations, a signal physically linked to localized land use, soil properties, and hydro-climatic conditions. Model predictability was highest for SPI-3, consistent with the region's dominant monsoon cycle, while lower SPI-6 performance highlights the limitations of a univariate approach in the presence of large-scale drivers such as the El Niño–Southern Oscillation (ENSO). Forecasts for 2025–2026 reveal three distinct risk profiles: a consistent drying trend at one station, increased rainfall volatility at another, and relative stability at the third. These findings demonstrate that effective drought management in complex equatorial regions requires localized, station-specific early warning systems rather than uniform city-wide approaches.]]>
