<![CDATA[Air quality forecasting in maritime tropical regions is challenged by zero-inflated rainfall regimes, where prolonged dry periods are intermittently disrupted by extreme precipitation, generating highly non-linear PM₂.₅ dynamics and limiting the effectiveness of conventional predictive models. This study evaluate the predictive performance of Recurrent Neural Networks (RNN) and Gated Recurrent Units (GRU) under such distributional conditions. A quantitative experimental design with a comparative approach is employed using 1,461 daily observations from the Central Java Climatology Station, incorporating rainfall, temperature, and relative humidity as predictors; a chronological data split preserves temporal dependencies, and performance is assessed using MAE, RMSE, MAPE, and R² metrics. The results indicate that GRU achieves only a marginal 4.3% reduction in MAE relative to RNN, while both models exhibit substantial predictive failure, as evidenced by negative R² values and MAPE exceeding 300%, with predictions collapsing toward the mean and failing to capture extreme pollution events. These findings demonstrate that standard recurrent architectures with conventional loss functions are intrinsically limited in modeling zero-inflated environmental data, contributing empirical evidence on the boundary conditions of deep learning in tropical air quality forecasting and underscoring the necessity for specialized modeling approaches to support reliable early warning systems.]]>
