<![CDATA[Fine particulate matter (PM2.5) pollution in Jakarta reached an annual average of 37.3 µg/m³ in 2023, 7.4 times the WHO threshold, causing over 10,000 premature deaths annually. Accurate short-term prediction is essential for early-warning systems, yet two gaps remain common in decomposition-based deep learning literature: Variational Mode Decomposition (VMD) parameters are rarely selected via systematic sensitivity analysis, and prediction bias is rarely corrected explicitly. This study addresses both gaps using 32,144 PM2.5 observations from Jakarta (2015-2025). A grid search over 20 parameter combinations identified K=6, alpha=500 as optimal (reconstruction error 1.88%). Each of six IMFs was predicted independently using CNN1D, reconstructed additively, and bias-corrected using validation-set mean bias error. Decomposition, not model complexity, drove accuracy: a non-decomposed baseline reached only R²=0.4302, versus R²=0.9151 (RMSE=4.83 µg/m³) for the proposed framework. We further validated fixed- and adaptive-parameter variants (VMD, AVMD) across 5 independent runs. VMD-CNN1D achieved R²=0.9177±0.0173, RMSE=4.7346±0.5064, while AVMD-CNN1D (K=7 selected consistently) achieved R²=0.9238±0.0089, RMSE=4.5388±0.2649. Diebold-Mariano tests showed AVMD outperforming VMD in 4 of 5 runs (p<0.0001) with lower variance, though a paired t-test across runs was not significant (p=0.684). Ablation identified the lowest-frequency IMF as most critical, consistent with Jakarta's dry-season and land-fire pollution patterns, while residual analysis revealed heteroscedasticity as a limitation. These findings show that rigorous parameter selection, bias correction, and multi-run validation, not architectural complexity alone, make decomposition-based deep learning reliable for PM2.5 prediction, offering a reproducibility-aware baseline for a future Jakarta air-quality early-warning system.]]>
