<![CDATA[Air pollution has become a critical environmental issue, particularly in urban areas such as DKI Jakarta, where pollutant concentrations frequently reach the highest levels in Indonesia. Accurate prediction of the Air Pollution Standard Index (ISPU) is essential for mitigating the adverse health and environmental impacts of poor air quality. However, ISPU data exhibit nonlinear, volatile, and non-stationary characteristics, posing challenges for conventional prediction models. To overcome these challenges, this study proposes a hybrid Complete Ensemble Empirical Mode Decomposition with Adaptive Noise–Long Short-Term Memory (CEEMDAN–LSTM) model, applied to daily ISPU data from 2010 to 2025 comprising 5,686 records. CEEMDAN was selected over conventional decomposition methods such as EEMD and VMD due to its ability to suppress mode-mixing and extract more stable Intrinsic Mode Functions (IMFs) through adaptive noise addition, thereby enhancing signal interpretability and learning efficiency. The ISPU time series was decomposed into multiple IMFs, and the resulting components were reconstructed and modeled using an optimized LSTM architecture obtained through Bayesian hyperparameter tuning. The optimal configuration batch size of 54, dropout rate of 0.37, and hidden units of 6, 33, and 34 achieved an RMSE of 14.0, reflecting a substantial improvement over the baseline LSTM model. The results demonstrate that integrating CEEMDAN with LSTM effectively reduces signal complexity, stabilizes convergence, and improves forecasting accuracy for non-stationary air quality data in DKI Jakarta. This modeling framework provides a robust foundation for developing predictive early-warning systems, supporting evidence-based environmental policy, and enhancing public health preparedness in rapidly urbanizing regions.]]>
